{ "cells": [ { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "Gabriele Iocco\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Analisi e Classificazione delle Email per la Rilevazione di SPAM

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Integrazione e controllo dell'hardware CUDA per Machine Learning

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Gestione della Memoria GPU e Allocazioni Dinamiche

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Dispositivi GPU rilevati: [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]\nCUDA available: True\nCUDA device count: 1\nCUDA device name: NVIDIA GeForce RTX 4070 Laptop GPU\nDispositivo corrente: 0\nTensorFlow version: 2.18.0\nCUDA disponibile: [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]\ncuDNN Version: 9\nCUDA Version: 12.5.1\nTest CUDA: True\nGPU available: True\n
\n
\n
\n
\n
\n
\n
I0000 00:00:1737558820.769880   19448 gpu_device.cc:2022] Created device /device:GPU:0 with 5472 MB memory:  -> device: 0, name: NVIDIA GeForce RTX 4070 Laptop GPU, pci bus id: 0000:01:00.0, compute capability: 8.9\n
\n
\n
\n
\n
" ] } } ], "source": [ "import warnings\n", "warnings.filterwarnings(\"ignore\")\n", "\n", "import os\n", "os.environ['TF_CPP_MIN_LOG_LEVEL']='1'\n", "\n", "import tensorflow as tf\n", "tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)\n", "\n", "gpu_devices = tf.config.list_physical_devices('GPU')\n", "\n", "if gpu_devices:\n", " try:\n", " for gpu in gpu_devices:\n", " tf.config.experimental.set_memory_growth(gpu, True)\n", " except RuntimeError as e:\n", " print(e)\n", "\n", "print(\"Dispositivi GPU rilevati:\", gpu_devices)\n", "\n", "import torch\n", "print(f\"CUDA available: {torch.cuda.is_available()}\")\n", "print(f\"CUDA device count: {torch.cuda.device_count()}\")\n", "print(f\"CUDA device name: {torch.cuda.get_device_name(0)}\" if torch.cuda.is_available() else \"No CUDA device found\")\n", "print(\"Dispositivo corrente:\", torch.cuda.current_device() if torch.cuda.is_available() else \"CPU\")\n", "print(\"TensorFlow version:\", tf.__version__)\n", "print(\"CUDA disponibile:\", tf.config.list_physical_devices('GPU'))\n", "print(\"cuDNN Version:\", tf.sysconfig.get_build_info()['cudnn_version'])\n", "print(\"CUDA Version:\", tf.sysconfig.get_build_info()[\"cuda_version\"])\n", "print(\"Test CUDA:\", tf.test.is_built_with_cuda())\n", "print(\"GPU available:\", tf.test.is_gpu_available())\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Configurazioni GPU TensorFlow\n", "os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'\n", "os.environ['XLA_FLAGS'] = '--xla_gpu_cuda_data_dir=/usr/lib/cuda'\n", "os.environ['TF_GPU_ALLOCATOR'] = 'cuda_malloc_async'\n", "os.environ['CUDA_VISIBLE_DEVICES'] = '0'\n", "os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'\n", "\n", "# Configurazione della memoria GPU\n", "gpus = tf.config.list_physical_devices('GPU')\n", "if gpus:\n", " for gpu in gpus:\n", " try:\n", " tf.config.experimental.set_memory_growth(gpu, True)\n", " except RuntimeError as e:\n", " print(e)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'\n", "os.environ['XLA_FLAGS'] = '--xla_gpu_cuda_data_dir=/usr/lib/cuda'\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
CuPy version: 13.3.0\n
\n
\n
\n
\n
" ] } } ], "source": [ "import cupy as cp\n", "print(\"CuPy version:\", cp.__version__)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Preferenza GPU da SpaCy: True\n
\n
\n
\n
\n
" ] } } ], "source": [ "from thinc.api import require_gpu, set_gpu_allocator\n", "import spacy\n", "\n", "require_gpu() # Forza l'uso della GPU\n", "nlp = spacy.load(\"en_core_web_trf\")\n", "print(\"Preferenza GPU da SpaCy:\", spacy.prefer_gpu())\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Analisi del dataset

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from colorama import Fore, Back, Style\n", "from tensorflow.keras.backend import clear_session\n", "import pandas as pd\n", "import matplotlib.pyplot as plt\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Funzione personalizzata per colorare il testo\n", "def print_colored(text, color=\"white\", bg_color=None, end=\"\\n\"):\n", " # Dizionario dei colori del testo\n", " color_dict = {\n", " 'red': Fore.RED,\n", " 'blue': Fore.BLUE,\n", " 'white': '\\033[97m', # Bianco puro (ANSI)\n", " 'black': Fore.BLACK\n", " }\n", "\n", " # Dizionario dei colori dello sfondo\n", " bg_color_dict = {\n", " 'black': Back.BLACK,\n", " 'blue': Back.BLUE,\n", " 'white': Back.WHITE\n", " }\n", "\n", " color_code = color_dict.get(color.lower(), '\\033[97m') \n", " bg_color_code = bg_color_dict.get(bg_color.lower(), '') if bg_color else ''\n", " \n", " print(f\"{color_code}{bg_color_code}{text}{Style.RESET_ALL}\", end=end)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "BASE_URL=\"/home/gap/Scrivania/Analisi_spam/\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "df_email = pd.read_csv(BASE_URL + \"spam_dataset.csv\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Il peso del dataset in memoria è di:\n5.66 MB\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Calcolo del peso totale del dataset in memoria\n", "dataset_size_ham_spam = df_email.memory_usage(deep=True).sum()\n", "\n", "# Conversione in megabyte (MB)\n", "dataset_size_ham_spam_mb = dataset_size_ham_spam / (1024 ** 2)\n", "\n", "print_colored(f\"Il peso del dataset in memoria è di:\", \"blue\") \n", "print(f\"{dataset_size_ham_spam_mb:.2f} MB\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
Out[12]:
\n
\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Unnamed: 0labeltextlabel_num
0605hamSubject: enron methanol ; meter # : 988291\\nth...0
12349hamSubject: hpl nom for january 9 , 2001\\n( see a...0
23624hamSubject: neon retreat\\nho ho ho , we ' re arou...0
34685spamSubject: photoshop , windows , office . cheap ...1
42030hamSubject: re : indian springs\\nthis deal is to ...0
\n
\n
\n
\n
\n
" ] } } ], "source": [ "df_email.head()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
<class 'pandas.core.frame.DataFrame'>\nRangeIndex: 5171 entries, 0 to 5170\nData columns (total 4 columns):\n #   Column      Non-Null Count  Dtype \n---  ------      --------------  ----- \n 0   Unnamed: 0  5171 non-null   int64 \n 1   label       5171 non-null   object\n 2   text        5171 non-null   object\n 3   label_num   5171 non-null   int64 \ndtypes: int64(2), object(2)\nmemory usage: 161.7+ KB\n
\n
\n
\n
\n
" ] } } ], "source": [ "df_email.info()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Numero di righe:\n5171\n\nNumero di colonne:\n4\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Numero di righe:\", \"blue\") \n", "print(df_email.shape[0])\n", "print()\n", "print_colored(\"Numero di colonne:\", \"blue\") \n", "print(df_email.shape[1])\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
Out[15]:
\n
\n
Unnamed: 0    5171\nlabel         5171\ntext          5171\nlabel_num     5171\ndtype: int64
\n
\n
\n
\n
" ] } } ], "source": [ "df_email.count()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Valori totali presenti nel dataset:\n20684\nValori totali 'non_missing' presenti nel dataset:\n20684\n\nValori mancanti\n0\n
\n
\n
\n
\n
" ] } } ], "source": [ "total_values = df_email.size\n", "print_colored(\"Valori totali presenti nel dataset:\", \"blue\")\n", "print(total_values)\n", "\n", "non_missing_values = df_email.count().sum()\n", "print_colored(\"Valori totali 'non_missing' presenti nel dataset:\", \"blue\")\n", "print(non_missing_values)\n", "\n", "missing_values = total_values - non_missing_values\n", "print_colored(\"\\nValori mancanti\", \"blue\")\n", "print(missing_values)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Valori mancanti per colonna:\n\n      Unnamed: 0  label   text  label_num\n0          False  False  False      False\n1          False  False  False      False\n2          False  False  False      False\n3          False  False  False      False\n4          False  False  False      False\n...          ...    ...    ...        ...\n5166       False  False  False      False\n5167       False  False  False      False\n5168       False  False  False      False\n5169       False  False  False      False\n5170       False  False  False      False\n\n[5171 rows x 4 columns]\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Valori mancanti per colonna:\\n\", \"blue\")\n", "print(df_email.isna())\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Valori mancanti per colonna:\n\n      Unnamed: 0  label   text  label_num\n0          False  False  False      False\n1          False  False  False      False\n2          False  False  False      False\n3          False  False  False      False\n4          False  False  False      False\n...          ...    ...    ...        ...\n5166       False  False  False      False\n5167       False  False  False      False\n5168       False  False  False      False\n5169       False  False  False      False\n5170       False  False  False      False\n\n[5171 rows x 4 columns]\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Valori mancanti per colonna:\\n\", \"blue\")\n", "print(df_email.isnull())\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
Out[19]:
\n
\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Unnamed: 0label_num
count5171.0000005171.000000
mean2585.0000000.289886
std1492.8834520.453753
min0.0000000.000000
25%1292.5000000.000000
50%2585.0000000.000000
75%3877.5000001.000000
max5170.0000001.000000
\n
\n
\n
\n
\n
" ] } } ], "source": [ "df_email.describe()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Tutte le righe corrispondono correttamente\nham -> 0, spam -> 1\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Controllo di possibili incongruenze tra label e label_num\n", "# La variabile inconsistencies viene popolata con le righe che non \n", "# rispettano le corrispondenze attese\n", "inconsistencies = df_email[\n", " ((df_email['label'] == 'ham') & (df_email['label_num'] != 0)) | \n", " ((df_email['label'] == 'spam') & (df_email['label_num'] != 1))\n", "]\n", "\n", "if inconsistencies.empty:\n", " print_colored(\"Tutte le righe corrispondono correttamente\", \"blue\")\n", " print (\"ham -> 0, spam -> 1\")\n", "else:\n", " print_colored(f\"Ci sono {len(inconsistencies)} righe con incongruenze:\", \"blue\")\n", " print(inconsistencies)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Visto che c'è una perfetta corrispondenza tra le feature \"label\" e \"label_num\", faccio il drop della colonna \"label\", anche perché, se scegliessi la feature categorica, dovrei applicare la tecnica del Label Encoding per convertire le variabili categoriali in formato numerico.

\n

Faccio il drop anche della colonna \"Unnamed:0\".
La colonna Unnamed:0 appare frequentemente quando si caricano file CSV in un DataFrame di pandas.

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Dataset dropped\n\n                                                text  label_num\n0  Subject: enron methanol ; meter # : 988291\\nth...          0\n1  Subject: hpl nom for january 9 , 2001\\n( see a...          0\n2  Subject: neon retreat\\nho ho ho , we ' re arou...          0\n3  Subject: photoshop , windows , office . cheap ...          1\n4  Subject: re : indian springs\\nthis deal is to ...          0\n
\n
\n
\n
\n
" ] } } ], "source": [ "df_email_dropped = df_email.drop(columns=['Unnamed: 0', 'label'])\n", "\n", "print_colored(\"Dataset dropped\\n\", \"blue\")\n", "print(df_email_dropped.head())\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Verifico la distribuzione di ham e spam nel dataset droppato

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "label_count=df_email_dropped[df_email_dropped.columns[1]].value_counts()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Distribuzione di ham e spam\n\nlabel_num\n0    3672\n1    1499\nName: count, dtype: int64\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Distribuzione di ham e spam\\n\", \"blue\")\n", "\n", "print(label_count)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "label_count.plot(kind='bar', figsize=(10, 6))\n", "plt.title(\"\\nDistribuzione di ham e spam\", fontsize=18)\n", "plt.xlabel(\"Label (0 = ham, 1 = spam)\", fontsize=14)\n", "plt.ylabel(\"Conteggio\", fontsize=14)\n", "plt.xticks(fontsize=12, color='#b81414', rotation=0)\n", "plt.yticks(fontsize=12, color='#b81414')\n", "plt.grid(axis='y', linestyle='--', alpha=0.5, color=\"#1f77b4\")\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
La differenza in percentuale tra ham e spam è del:\n42.02%\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Calcolo della differenza in percentuale\n", "total_label = label_count.sum()\n", "difference_percentage = ((label_count.max() - label_count.min()) / total_label) * 100\n", "\n", "print_colored(f\"La differenza in percentuale tra ham e spam è del:\", \"blue\") \n", "print(f\"{difference_percentage:.2f}%\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Preprocessing del Testo delle email SPAM

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
[nltk_data] Downloading package stopwords to /home/gap/nltk_data...\n[nltk_data]   Package stopwords is already up-to-date!\n[nltk_data] Downloading package wordnet to /home/gap/nltk_data...\n[nltk_data]   Package wordnet is already up-to-date!\n[nltk_data] Downloading package omw-1.4 to /home/gap/nltk_data...\n[nltk_data]   Package omw-1.4 is already up-to-date!\n
\n
\n
\n
\n
Out[26]:
\n
\n
True
\n
\n
\n
\n
" ] } } ], "source": [ "import re\n", "import nltk\n", "import nltk.corpus\n", "from nltk.corpus import stopwords\n", "from nltk.stem import WordNetLemmatizer\n", "from nltk import download\n", "\n", "download('stopwords')\n", "download('wordnet')\n", "download('omw-1.4')\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Dataset 'df_email_dropped_spam' senza l'applicazione della funzione 'clean_text' alla colonna 'text'\n                                                 text\n3   Subject: photoshop , windows , office . cheap ...\n7   Subject: looking for medication ? we ` re the ...\n10  Subject: vocable % rnd - word asceticism\\nvcsc...\n11  Subject: report 01405 !\\nwffur attion brom est...\n13  Subject: vic . odin n ^ ow\\nberne hotbox carna...\n\nDataset 'df_email_dropped_spam' con l'applicazione della funzione 'clean_text' alla colonna 'text'\n                                         cleaned_text\n3   photoshop windows office cheap main trend abas...\n7   look medication best source difficult make mat...\n10  vocable rnd word asceticism vcsc brand new sto...\n11  report wffur attion brom est inst siupied pgst...\n13  vic odin n ow berne hotbox carnal bride cutwor...\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Inizializzazione di stopwords e lemmatizer\n", "stop_words = set(stopwords.words('english')) # Stopword per la lingua inglese\n", " # Rimuove parole comuni inglesi per ridurre il rumore nei dati\n", "lemmatizer = WordNetLemmatizer() # Riduce le parole alla loro forma base per \n", " # diminuire la dimensionalità del vocabolario\n", "\n", "def clean_text(text):\n", " # Rimozione di caratteri speciali e punteggiatura\n", " text = re.sub(r'[^a-zA-Z\\s]', '', text) \n", " \n", " # Conversione di tutte le lettere in minuscolo\n", " text = text.lower()\n", " \n", " # Rimozione della parola 'subject' se presente perché si riferisce alla struttura standard di un'email\n", " # non aggiunge valore informativo per l'analisi, anzi, può influenzarla negativamente\n", " text = text.replace('subject', '') \n", " \n", " tokens = [lemmatizer.lemmatize(word, pos='v') for word in text.split() if word not in stop_words]\n", " return ' '.join(tokens)\n", "\n", "# Filtro le email SPAM e creo una copia per evitare SettingWithCopyWarning\n", "# Il warning SettingWithCopyWarning si verifica quando si modificano direttamente valori di un DataFrame filtrato\n", "df_email_dropped_spam = df_email_dropped[df_email_dropped['label_num'] == 1].copy()\n", "\n", "# Creazione della colonna cleaned_text con i dati della colonna text puliti\n", "df_email_dropped_spam['cleaned_text'] = df_email_dropped_spam['text'].apply(clean_text)\n", "\n", "print_colored(\"Dataset 'df_email_dropped_spam' senza l'applicazione della funzione 'clean_text' alla colonna 'text'\", \"blue\")\n", "print(df_email_dropped_spam[['text']].head())\n", "\n", "print()\n", "\n", "print_colored(\"Dataset 'df_email_dropped_spam' con l'applicazione della funzione 'clean_text' alla colonna 'text'\", \"blue\")\n", "print(df_email_dropped_spam[['cleaned_text']].head())\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
                                                 text  label_num  \\\n3   Subject: photoshop , windows , office . cheap ...          1   \n7   Subject: looking for medication ? we ` re the ...          1   \n10  Subject: vocable % rnd - word asceticism\\nvcsc...          1   \n11  Subject: report 01405 !\\nwffur attion brom est...          1   \n13  Subject: vic . odin n ^ ow\\nberne hotbox carna...          1   \n\n                                         cleaned_text  \n3   photoshop windows office cheap main trend abas...  \n7   look medication best source difficult make mat...  \n10  vocable rnd word asceticism vcsc brand new sto...  \n11  report wffur attion brom est inst siupied pgst...  \n13  vic odin n ow berne hotbox carnal bride cutwor...  \n
\n
\n
\n
\n
" ] } } ], "source": [ "print(df_email_dropped_spam.head())\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
Out[29]:
\n
\n
label_num\n1    1499\nName: count, dtype: int64
\n
\n
\n
\n
" ] } } ], "source": [ "# Verifico se il conteggio dei valori della colonna 'label num' riferito \n", "# alle mail spam sia uguale al conteggio di partenza\n", "df_email_dropped_spam[df_email_dropped_spam.columns[1]].value_counts()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Il peso del dataset droppato, pulito e contenente solo spam in memoria è di:\n1.28 MB\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Calcolo del peso totale del dataset contenente solo SPAM droppato e pulito in memoria\n", "dataset_size_dropped_spam = df_email_dropped_spam[['cleaned_text', 'label_num']].memory_usage(deep=True).sum()\n", "\n", "# Conversione in megabyte (MB)\n", "dataset_size_mb_dropped = dataset_size_dropped_spam / (1024 ** 2)\n", "\n", "print_colored(f\"Il peso del dataset droppato, pulito e contenente solo spam in memoria è di:\", \"blue\") \n", "print(f\"{dataset_size_mb_dropped:.2f} MB\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
La memoria liberata grazie al dropout e alla pulizia del testo è di:\n4.38 MB\nPercentuale di memoria liberata:\n63.10 %\n
\n
\n
\n
\n
" ] } } ], "source": [ "dataset_size_difference = dataset_size_ham_spam_mb - dataset_size_mb_dropped\n", "\n", "print_colored(f\"La memoria liberata grazie al dropout e alla pulizia del testo è di:\", \"blue\") \n", "print(f\"{dataset_size_difference:.2f} MB\")\n", "\n", "total_size = dataset_size_ham_spam_mb + dataset_size_mb_dropped\n", "difference_percentage_size = (dataset_size_difference/total_size)*100\n", "\n", "print_colored(f\"Percentuale di memoria liberata:\", \"blue\") \n", "print(f\"{difference_percentage_size:.2f} %\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Il numero totale di parole nella colonna 'cleaned_text' è di:\n177805\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Calcolo del numero totale di parole nella colonna 'cleaned_text'\n", "total_word_count = df_email_dropped_spam['cleaned_text'].str.split().str.len().sum()\n", "\n", "print_colored(f\"Il numero totale di parole nella colonna 'cleaned_text' è di:\", \"blue\")\n", "print(f\"{total_word_count}\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Implementazione Multi-Modello

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Costruzione e addestramento di modelli Word2Vec per l'analisi delle email classificate come SPAM

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Utilizzerò la tecnica Word2Vec per rappresentare le parole come vettori densi, ovvero rappresentazioni numeriche compatte che collocano i dati nello spazio semantico. Questo metodo consente di ottenere embedding capaci di catturare le relazioni semantiche tra le parole, posizionandole in uno spazio vettoriale in cui parole dal significato simile risultano vicine tra loro. Successivamente, applicherò l'algoritmo K-Means per il clustering, al fine di identificare gruppi semantici significativi generati dal modello. Questo approccio risulta particolarmente efficace in contesti caratterizzati da dati rumorosi o ad alta dimensionalità semantica.

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Gli obiettivi di questi modelli saranno:

- Individuare i Topic principali tra le email classificate come SPAM.

- Calcolare la distanza semantica tra i topic ottenuti per valutare l'eterogeneità dei contenuti delle email SPAM.

- Analisi contenutistica approfondita individuando i principali topic trattati nelle email SPAM consentendo di ottenere informazioni preziose sui trend, tematiche e schemi ricorrenti, potenziando le strategie di cybersecurity.

- Valutazione dell'eterogeneità calcolando la distanza semantica tra i topic e quindi comprendere la diversità dei contenuti SPAM, utile per ottimizzare le difese contro un'ampia gamma di attacchi.

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Addestramento del modello: model_vec50_win3\nAddestramento del modello: model_vec100_win3\nAddestramento del modello: model_vec100_win5\nAddestramento del modello: model_vec200_win3\nAddestramento del modello: model_vec200_win5\n
\n
\n
\n
\n
" ] } } ], "source": [ "from gensim.models import Word2Vec\n", "from gensim.models import CoherenceModel\n", "from gensim.corpora.dictionary import Dictionary\n", "from sklearn.metrics import silhouette_score\n", "from sklearn.cluster import KMeans\n", "import multiprocessing\n", "\n", "# Tokenizzazione\n", "# divide il testo in token (parole e sottoparole) e assegna un identificatore numerico univoco a ogni token\n", "tokenized_texts = [text.split() for text in df_email_dropped_spam['cleaned_text']]\n", "\n", "# Creazione del dizionario per Coherence Score\n", "dictionary = Dictionary(tokenized_texts)\n", "\n", "# Configurazioni da testare\n", "configs = [\n", " {\"name\": \"model_vec50_win3\", \"vector_size\": 50, \"window\": 3, \"min_count\": 2, \"sg\": 1},\n", " {\"name\": \"model_vec100_win3\", \"vector_size\": 100, \"window\": 3, \"min_count\": 2, \"sg\": 1},\n", " {\"name\": \"model_vec100_win5\", \"vector_size\": 100, \"window\": 5, \"min_count\": 2, \"sg\": 1},\n", " {\"name\": \"model_vec200_win3\", \"vector_size\": 200, \"window\": 3, \"min_count\": 2, \"sg\": 1},\n", " {\"name\": \"model_vec200_win5\", \"vector_size\": 200, \"window\": 5, \"min_count\": 2, \"sg\": 1},\n", "]\n", "\n", "# Funzione per estrarre i topic da Word2Vec\n", "def extract_topics_from_word2vec(model, n_topics, topn=10):\n", " word_vectors = np.array([model.wv[word] for word in model.wv.index_to_key])\n", " words = model.wv.index_to_key\n", " \n", " # Clustering con KMeans\n", " kmeans = KMeans(n_clusters=n_topics, random_state=42)\n", " labels = kmeans.fit_predict(word_vectors)\n", " \n", " # Raggruppamento delle parole per cluster\n", " topics = []\n", " for cluster in range(n_topics):\n", " words_in_cluster = [words[i] for i, label in enumerate(labels) if label == cluster]\n", " topics.append(words_in_cluster[:topn])\n", " return topics\n", "\n", "# Per salvare i risultati\n", "results_main = []\n", "\n", "# Numero di topic da testare\n", "n_topics_list = [2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20]\n", "\n", "# Addestramento e valutazione dei modelli\n", "for config in configs:\n", " print_colored(f\"Addestramento del modello: {config['name']}\", \"blue\")\n", " \n", " # Addestramento del modello Word2Vec\n", " \"\"\"\n", " vector_size indica la dimensione del vettore generato per ogni parola\n", " window indica la dimensione del contesto, il numero di parole vicine considerate per ogni parola\n", " Siccome il progetto prevede un'analisi contenutistica approfondita lo imposto a 3 e 5\n", " min_count, le parole con una frequenza inferiore a questo valore vengono ignorate\n", " sg Skip-gram indica il tipo di architettura\n", " Skip-gram cattura relazioni semantiche\n", " \"\"\"\n", "\n", " # Embedding delle parole\n", " # converte i token in vettori densi, cattura relazioni semantiche tra token\n", " # rappresenta il significato contestuale dei token\n", " model = Word2Vec(\n", " sentences=tokenized_texts,\n", " vector_size=config[\"vector_size\"],\n", " window=config[\"window\"],\n", " min_count=config[\"min_count\"],\n", " workers=multiprocessing.cpu_count(),\n", " sg=config[\"sg\"]\n", " )\n", " model.save(f\"{config['name']}.model\")\n", " \n", " for n_topics in n_topics_list:\n", " try:\n", " # Estrazione dei topic\n", " topics = extract_topics_from_word2vec(model, n_topics)\n", "\n", " # Calcolo del Coherence Score\n", " coherence_model = CoherenceModel(\n", " topics=topics,\n", " texts=tokenized_texts,\n", " dictionary=dictionary,\n", " coherence=\"c_v\"\n", " )\n", " coherence_score = coherence_model.get_coherence()\n", "\n", " # Calcolo del Silhouette Score\n", " word_vectors = np.array([model.wv[word] for word in model.wv.index_to_key])\n", " silhouette_avg = silhouette_score(word_vectors, KMeans(n_clusters=n_topics, random_state=42).fit_predict(word_vectors))\n", "\n", " # Salvataggio dei risultati\n", " results_main.append({\n", " \"model_name\": config[\"name\"],\n", " \"vector_size\": config[\"vector_size\"],\n", " \"window\": config[\"window\"],\n", " \"n_topics\": n_topics,\n", " \"coherence_c_v\": coherence_score,\n", " \"silhouette\": silhouette_avg,\n", " })\n", "\n", " except Exception as e:\n", " print(f\"Errore per {config['name']} con {n_topics} topic: {e}\")\n", " results_main.append({\n", " \"model_name\": config[\"name\"],\n", " \"vector_size\": config[\"vector_size\"],\n", " \"window\": config[\"window\"],\n", " \"n_topics\": n_topics,\n", " \"coherence_c_v\": None,\n", " \"silhouette\": None,\n", " })\n", "\n", "# Conversione in DataFrame per analisi\n", "results_df_main = pd.DataFrame(results_main)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Modelli costruiti e addestrati:\nmodel_vec50_win3\nmodel_vec100_win3\nmodel_vec100_win5\nmodel_vec200_win3\nmodel_vec200_win5\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(f\"Modelli costruiti e addestrati:\", \"blue\")\n", "\n", "for config in configs: \n", " print(config['name'])\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Colonne disponibili:\nIndex(['model_name', 'vector_size', 'window', 'n_topics', 'coherence_c_v',\n       'silhouette'],\n      dtype='object')\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Colonne disponibili:\", \"blue\")\n", "print(results_df_main.columns)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nRisultati del Coherence Score (C_v) con evidenziazione dei valori massimi:\n╒════════════╤═════════════════════╤═════════════════════╤═════════════════════╤═════════════════════╤════════════════════╕\n│   n_topics │   model_vec100_win3 │   model_vec100_win5 │   model_vec200_win3 │   model_vec200_win5 │   model_vec50_win3 │\n╞════════════╪═════════════════════╪═════════════════════╪═════════════════════╪═════════════════════╪════════════════════╡\n│          2 │              0.7562 │              0.7411 │              0.7569 │              0.7639 │             0.7638 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          3 │              0.7526 │              0.5601 │              0.6453 │              0.5375 │             0.7608 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          4 │              0.667  │              0.6081 │              0.5204 │              0.6148 │             0.6396 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          5 │              0.578  │              0.6189 │              0.5869 │              0.626  │             0.5702 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          6 │              0.6531 │              0.589  │              0.6575 │              0.6168 │             0.6237 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          7 │              0.6303 │              0.6398 │              0.6746 │              0.6482 │             0.6777 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          8 │              0.6281 │              0.6586 │              0.6492 │              0.612  │             0.6283 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          9 │              0.6607 │              0.6629 │              0.6947 │              0.6441 │             0.6354 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│         10 │              0.6214 │              0.6308 │              0.6724 │              0.6038 │             0.6367 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│         15 │              0.6152 │              0.6355 │              0.6521 │              0.6793 │             0.6612 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│         20 │              0.6552 │              0.6582 │              0.6795 │              0.6518 │             0.6471 │\n╘════════════╧═════════════════════╧═════════════════════╧═════════════════════╧═════════════════════╧════════════════════╛\n
\n
\n
\n
\n
" ] } } ], "source": [ "from tabulate import tabulate\n", "\n", "# Raggruppamento per modello e numero di topic\n", "table = results_df_main.pivot(\n", " index=\"n_topics\", \n", " columns=\"model_name\", \n", " values=\"coherence_c_v\"\n", ")\n", "\n", "# Riordino le colonne per chiarezza\n", "table = table[sorted(table.columns)]\n", "\n", "# Evidenziazione dei valori più alti in ogni riga\n", "def highlight_max(table):\n", " \"\"\"\n", " Evidenzia il valore massimo di ogni riga.\n", " \"\"\"\n", " highlighted_table = []\n", " for row in table.itertuples(index=True):\n", " row_list = [row.Index] # Includo il numero di topic come prima colonna\n", " max_value = max(row[1:]) # Trovo il massimo nella riga\n", " for value in row[1:]:\n", " if value == max_value:\n", " row_list.append(f\"\\033[1;34m{value:.4f}\\033[0m\") # Evidenzia in blu\n", " else:\n", " row_list.append(f\"{value:.4f}\")\n", " highlighted_table.append(row_list)\n", " return highlighted_table\n", "\n", "# Applico l'evidenziazione\n", "highlighted_table = highlight_max(table)\n", "\n", "# Preparo l'intestazione della tabella\n", "headers = [\"n_topics\"] + list(table.columns)\n", "\n", "# Colore blu per gli header\n", "blue_headers = [f\"\\033[94m{header}\\033[0m\" for header in headers]\n", "\n", "print_colored(\"\\nRisultati del Coherence Score (C_v) con evidenziazione dei valori massimi:\", \"blue\")\n", "print(tabulate(highlighted_table, headers=blue_headers, tablefmt=\"fancy_grid\"))\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nRisultati della metrica Silhouette con evidenziazione dei valori massimi:\n╒════════════╤═════════════════════╤═════════════════════╤═════════════════════╤═════════════════════╤════════════════════╕\n│   n_topics │   model_vec100_win3 │   model_vec100_win5 │   model_vec200_win3 │   model_vec200_win5 │   model_vec50_win3 │\n╞════════════╪═════════════════════╪═════════════════════╪═════════════════════╪═════════════════════╪════════════════════╡\n│          2 │              0.6166 │              0.6016 │              0.6213 │              0.6031 │             0.6056 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          3 │              0.6031 │              0.5912 │              0.558  │              0.5973 │             0.5863 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          4 │              0.5216 │              0.5134 │              0.5339 │              0.5039 │             0.4949 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          5 │              0.4931 │              0.3755 │              0.5074 │              0.3824 │             0.478  │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          6 │              0.435  │              0.3449 │              0.4417 │              0.3493 │             0.4155 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          7 │              0.4413 │              0.3469 │              0.4266 │              0.3515 │             0.426  │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          8 │              0.4524 │              0.3441 │              0.3666 │              0.3457 │             0.3102 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│          9 │              0.4528 │              0.3309 │              0.3695 │              0.3438 │             0.3121 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│         10 │              0.3494 │              0.3008 │              0.3657 │              0.364  │             0.3194 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│         15 │              0.2904 │              0.2217 │              0.3222 │              0.2783 │             0.2588 │\n├────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼────────────────────┤\n│         20 │              0.2915 │              0.2068 │              0.3219 │              0.2215 │             0.2134 │\n╘════════════╧═════════════════════╧═════════════════════╧═════════════════════╧═════════════════════╧════════════════════╛\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Raggruppamento per modello e numero di topic\n", "table = results_df_main.pivot(\n", " index=\"n_topics\", \n", " columns=\"model_name\", \n", " values=\"silhouette\"\n", ")\n", "\n", "# Riordino le colonne per chiarezza\n", "table = table[sorted(table.columns)]\n", "\n", "# Evidenziazione dei valori più alti in ogni riga\n", "def highlight_max(table):\n", " \"\"\"\n", " Evidenzia il valore massimo di ogni riga.\n", " \"\"\"\n", " highlighted_table = []\n", " for row in table.itertuples(index=True):\n", " row_list = [row.Index] # Numero di topic come prima colonna\n", " max_value = max(row[1:]) # Valore massimo nella riga\n", " for value in row[1:]:\n", " if value == max_value:\n", " row_list.append(f\"\\033[1;34m{value:.4f}\\033[0m\") # Evidenzia in blu\n", " else:\n", " row_list.append(f\"{value:.4f}\")\n", " highlighted_table.append(row_list)\n", " return highlighted_table\n", "\n", "# Applico l'evidenziazione\n", "highlighted_table = highlight_max(table)\n", "\n", "# Preparo l'intestazione della tabella\n", "headers = [\"n_topics\"] + list(table.columns)\n", "\n", "# Colore blu per gli header\n", "blue_headers = [f\"\\033[94m{header}\\033[0m\" for header in headers]\n", "\n", "print_colored(\"\\nRisultati della metrica Silhouette con evidenziazione dei valori massimi:\", \"blue\")\n", "print(tabulate(highlighted_table, headers=blue_headers, tablefmt=\"fancy_grid\"))\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Grafico del Coherence Score

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "plt.figure(figsize=(10, 6))\n", "for model_name in results_df_main[\"model_name\"].unique():\n", " \n", " model_data = results_df_main[results_df_main[\"model_name\"] == model_name]\n", " plt.plot(\n", " model_data[\"n_topics\"],\n", " model_data[\"coherence_c_v\"],\n", " marker=\"o\",\n", " label=model_name\n", " )\n", " \n", "plt.title(\"\\nCoherence Score (C_v) per Modello e Numero di Topic\", fontsize=18)\n", "plt.xlabel(\"Numero di Topic\", fontsize=14, color=\"#b81414\")\n", "plt.ylabel(\"Coherence Score (C_v)\", fontsize=14, color=\"#b81414\")\n", "plt.legend()\n", "plt.grid()\n", "plt.tight_layout()\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Grafico per la metrica Silhouette

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "plt.figure(figsize=(10, 6))\n", "for model_name in results_df_main[\"model_name\"].unique():\n", " \n", " model_data = results_df_main[results_df_main[\"model_name\"] == model_name]\n", " plt.plot(\n", " model_data[\"n_topics\"],\n", " model_data[\"silhouette\"],\n", " marker=\"o\",\n", " label=model_name\n", " )\n", "\n", "plt.title(\"\\nSilhouette Score per Modello e Numero di Topic\", fontsize=18)\n", "plt.xlabel(\"Numero di Topic\", fontsize=14, color=\"#b81414\")\n", "plt.ylabel(\"Silhouette Score\", fontsize=14, color=\"#b81414\")\n", "plt.legend()\n", "plt.grid()\n", "plt.tight_layout()\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nMigliori configurazioni per Coherence Score:\n╒════╤═══════════════════╤═══════════════════╤═══════════════════╕\n│    │ Modello           │   Numero di Topic │   Coherence Score │\n╞════╪═══════════════════╪═══════════════════╪═══════════════════╡\n│ 11 │ model_vec100_win3 │                 2 │          0.756224 │\n├────┼───────────────────┼───────────────────┼───────────────────┤\n│ 22 │ model_vec100_win5 │                 2 │          0.741138 │\n├────┼───────────────────┼───────────────────┼───────────────────┤\n│ 33 │ model_vec200_win3 │                 2 │          0.756923 │\n├────┼───────────────────┼───────────────────┼───────────────────┤\n│ 44 │ model_vec200_win5 │                 2 │          0.763862 │\n├────┼───────────────────┼───────────────────┼───────────────────┤\n│  0 │ model_vec50_win3  │                 2 │          0.76384  │\n╘════╧═══════════════════╧═══════════════════╧═══════════════════╛\n\nMigliori configurazioni per Silhouette Score:\n╒════╤═══════════════════╤═══════════════════╤════════════════════╕\n│    │ Modello           │   Numero di Topic │   Silhouette Score │\n╞════╪═══════════════════╪═══════════════════╪════════════════════╡\n│ 11 │ model_vec100_win3 │                 2 │           0.616605 │\n├────┼───────────────────┼───────────────────┼────────────────────┤\n│ 22 │ model_vec100_win5 │                 2 │           0.601632 │\n├────┼───────────────────┼───────────────────┼────────────────────┤\n│ 33 │ model_vec200_win3 │                 2 │           0.621298 │\n├────┼───────────────────┼───────────────────┼────────────────────┤\n│ 44 │ model_vec200_win5 │                 2 │           0.603138 │\n├────┼───────────────────┼───────────────────┼────────────────────┤\n│  0 │ model_vec50_win3  │                 2 │           0.605643 │\n╘════╧═══════════════════╧═══════════════════╧════════════════════╛\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Trovo i migliori risultati per ogni modello basati su Coherence Score\n", "best_coherence = results_df_main.loc[results_df_main.groupby(\"model_name\")[\"coherence_c_v\"].idxmax()]\n", "\n", "# Trovo i migliori risultati per ogni modello basati su Silhouette Score\n", "best_silhouette = results_df_main.loc[results_df_main.groupby(\"model_name\")[\"silhouette\"].idxmax()]\n", "\n", "# Formatto i risultati come tabelle\n", "coherence_table = best_coherence[[\"model_name\", \"n_topics\", \"coherence_c_v\"]]\n", "silhouette_table = best_silhouette[[\"model_name\", \"n_topics\", \"silhouette\"]]\n", "\n", "print_colored(\"\\nMigliori configurazioni per Coherence Score:\", \"blue\")\n", "print(tabulate(coherence_table, headers=[\"Modello\", \"Numero di Topic\", \"Coherence Score\"], tablefmt=\"fancy_grid\"))\n", "\n", "print_colored(\"\\nMigliori configurazioni per Silhouette Score:\", \"blue\")\n", "print(tabulate(silhouette_table, headers=[\"Modello\", \"Numero di Topic\", \"Silhouette Score\"], tablefmt=\"fancy_grid\"))\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Clustering con K-Means per ottenere i vettori delle parole

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Dopo aver addestrato uno o più modelli Word2Vec il passo successivo è quello di estrarre i vettori di tutte le parole dai modelli e stampare ad esempio le prime 10 parole del vocabolario per analizzare rapidamente un clustering specifico.

KMeans crea un clustering statico con un numero fisso di cluster scelto in modo arbitrario.

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nAnalisi per il modello: model_vec50_win3.model\nCaricamento del modello salvato: model_vec50_win3.model\nVocabolario: com, http, company, price, get, e, www, information, font, email\n\nNumero di Topic: 2\n  Topic 1: proletariat bronco cognitive kenton baseboard catherine metabole swamp demagnify cloudburst\n  Topic 2: com http company price get e www information font email\n\nNumero di Topic: 3\n  Topic 1: proletariat bronco cognitive baseboard catherine metabole swamp demagnify cloudburst buoy\n  Topic 2: company get information please statements stock time new may make\n  Topic 3: com http price e www font email td us nbsp\n\nNumero di Topic: 4\n  Topic 1: right top de oo health strong pain drive body weight\n  Topic 2: company get information email please statements us stock time new\n  Topic 3: com http price e www font td nbsp p height\n  Topic 4: ural tras ects reas rien expe wn vol ume eff\n\nNumero di Topic: 5\n  Topic 1: price get us new one go offer need free service\n  Topic 2: company information email please statements stock time may make use\n  Topic 3: com http e www font td nbsp p height size\n  Topic 4: top de oo strong drive age mr st city america\n  Topic 5: ural rien wn vol ume eff ger srge smallcap gs\n\nNumero di Topic: 6\n  Topic 1: service market inc business say international million high gas first\n  Topic 2: company information statements stock time may make use look within\n  Topic 3: com http www font td nbsp p height size width\n  Topic 4: top de leave head la ca female solid ce un\n  Topic 5: smallcap penls tier marketpiace eshopkey zwallet barreis chevron futureevents mw\n  Topic 6: price get e email please us new one go offer\n\nNumero di Topic: 7\n  Topic 1: service market inc business say products international million high gas\n  Topic 2: get e email please us new one use go offer\n  Topic 3: com http www font td nbsp p height size width\n  Topic 4: top de st leave die head usb wireless fat la\n  Topic 5: smallcap penls tier grecian marketpiace eshopkey jdz zwallet barreis chevron\n  Topic 6: price v windows microsoft software professional office xp adobe system\n  Topic 7: company information statements stock time may make look within report\n\nNumero di Topic: 8\n  Topic 1: nds spe mitosis quadrant eagle relaxation madras meson miliennium bra\n  Topic 2: br right top de oo strong drive body age st\n  Topic 3: injections cute doom rn slutty agaln thieve gall anthropomorphism ie\n  Topic 4: service market inc business say international million internet high call\n  Topic 5: font td nbsp p height size width b align tr\n  Topic 6: company information statements stock may make use look within report\n  Topic 7: com http e www email please message r send u\n  Topic 8: price get us time new one go offer need free\n\nNumero di Topic: 9\n  Topic 1: nds spe mitosis quadrant eagle relaxation madras meson miliennium bra\n  Topic 2: p br right top de oo strong body age st\n  Topic 3: injections cute doom rn slutty agaln thieve gall anthropomorphism ie\n  Topic 4: one service market inc business say products international million internet\n  Topic 5: price x v windows microsoft software professional office xp adobe\n  Topic 6: company information statements stock may make use look within report\n  Topic 7: com http e www email please us message r send\n  Topic 8: get time new go offer need free pills money take\n  Topic 9: font td nbsp height size width b align tr color\n\nNumero di Topic: 10\n  Topic 1: brendan sri nat brrating bd nds wor spe mitosis quadrant\n  Topic 2: top power age mr city lead design wait data music\n  Topic 3: injections cute doom rn slutty agaln thieve gall anthropomorphism ie\n  Topic 4: e p r b u c x v n th\n  Topic 5: price windows microsoft software professional office xp adobe system cd\n  Topic 6: company information statements stock may make look within report investment\n  Topic 7: com http www email please us message free send mail\n  Topic 8: one use go service market inc business account say news\n  Topic 9: font td nbsp height size width align tr color face\n  Topic 10: get time new offer need pills money take order want\n\nNumero di Topic: 15\n  Topic 1: bra ural tim tras ects reas saf rien expe kno\n  Topic 2: power control computer support age mr enter city case card\n  Topic 3: rn duve mexican crtc htmlbody wl tp boys internetstore zdrive\n  Topic 4: get us time one use go offer need free money\n  Topic 5: information stock within report investment securities provide act trade advice\n  Topic 6: statements may make look result future number forward could base\n  Topic 7: email please message send mail contact click remove account link\n  Topic 8: company service market include inc business say news share international\n  Topic 9: font td nbsp height size width align tr color face\n  Topic 10: new pills order online best save also available ship viagra\n  Topic 11: price windows microsoft software professional office xp adobe system cd\n  Topic 12: br top de oo strong drive body st id leave\n  Topic 13: e p r b u c x v n th\n  Topic 14: com http www computron net info src nd html index\n  Topic 15: fw beman compete beach respecter durrell influential california wound player\n\nNumero di Topic: 20\n  Topic 1: smallcap penls tier grecian marketpiace yasser eshopkey jdz zwallet barreis\n  Topic 2: increase well two program industry network hold credit control experience\n  Topic 3: rn crtc htmlbody zdrive goto poliomyelitis fingernail conceal macedonia kenney\n  Topic 4: right second power meet computer support enter type city case\n  Topic 5: windows microsoft professional office xp adobe cd ms pro mx\n  Topic 6: statements may look result future forward could base risk believe\n  Topic 7: email please message send mail contact click remove line list\n  Topic 8: time make one use go take want account like say\n  Topic 9: font td nbsp height width align tr border href src\n  Topic 10: price get us new offer need money order work online\n  Topic 11: software system download drive draw suite ibm pc creative dvd\n  Topic 12: top de oo st die digital head usb wireless fat\n  Topic 13: e p r b u c x v n th\n  Topic 14: size color face br family strong body leave rnd alt\n  Topic 15: nat lon congratulatory apr wow brrating satellite tri thu heal\n  Topic 16: information stock within report investment securities provide act trade advice\n  Topic 17: age mr data dr sport est jan sun europe africa\n  Topic 18: pills viagra cialis last soft hours per health mg pain\n  Topic 19: com http www free computron link net info stop html\n  Topic 20: company service market include inc business news share international million\n\nAnalisi per il modello: model_vec100_win3.model\nCaricamento del modello salvato: model_vec100_win3.model\nVocabolario: com, http, company, price, get, e, www, information, font, email\n\nNumero di Topic: 2\n  Topic 1: zero proletariat emotional bronco cognitive baseboard catherine bramble metabole swamp\n  Topic 2: com http company price get e www information font email\n\nNumero di Topic: 3\n  Topic 1: zero proletariat emotional bronco cognitive baseboard catherine bramble metabole swamp\n  Topic 2: company get information email please statements us stock time new\n  Topic 3: com http price e www font td nbsp p height\n\nNumero di Topic: 4\n  Topic 1: ural tim ects reas saf ire vol ume eff ger\n  Topic 2: company price get information email please statements us stock time\n  Topic 3: com http e www font td nbsp p height size\n  Topic 4: increase right top de oo health strong pain power computer\n\nNumero di Topic: 5\n  Topic 1: top de strong age mr st city lead leave data\n  Topic 2: company information email please statements stock time may make use\n  Topic 3: com http e www font td nbsp p height size\n  Topic 4: price get us new one go offer need free service\n  Topic 5: mirrior smallcap shaun conservative gs verizon nl penls scan johnson\n\nNumero di Topic: 6\n  Topic 1: top de leave la ca female solid ce und white\n  Topic 2: price get e email please us new one go offer\n  Topic 3: com http www font td nbsp p height size width\n  Topic 4: service market inc business say products international million high gas\n  Topic 5: smallcap penls walter amende reaily prlces eshopkey fu balustrade futureevents\n  Topic 6: company information statements stock time may make use look within\n\nNumero di Topic: 7\n  Topic 1: top de body leave fat la ca female solid ce\n  Topic 2: price get us new one offer need free pills money\n  Topic 3: http www font td nbsp p height size width b\n  Topic 4: go service market inc business say international million high call\n  Topic 5: smallcap penls walter amende reaily prlces eshopkey fu balustrade destroy\n  Topic 6: com e email please message r send mail contact click\n  Topic 7: company information statements stock time may make use look within\n\nNumero di Topic: 8\n  Topic 1: smallcap penls scan walter amende reaily prlces fa eshopkey fu\n  Topic 2: com http e www email please message free r send\n  Topic 3: get us time new one go offer need pills money\n  Topic 4: service market inc business say international million high br gas\n  Topic 5: price x v windows microsoft software professional office xp adobe\n  Topic 6: company information statements stock may make use look within report\n  Topic 7: font td nbsp p height size width b align tr\n  Topic 8: top de body st leave head fat la ca gb\n\nNumero di Topic: 9\n  Topic 1: smallcap penls scan walter amende reaily prlces fa eshopkey fu\n  Topic 2: email please message send mail contact click remove account link\n  Topic 3: get us time new one go offer need free pills\n  Topic 4: service market inc business say international million high gas first\n  Topic 5: price windows microsoft software professional office xp adobe system cd\n  Topic 6: company information statements stock may make use look within report\n  Topic 7: font td nbsp height size width align tr color face\n  Topic 8: top de body leave head fat la ca female solid\n  Topic 9: com http e www p r b u c computron\n\nNumero di Topic: 10\n  Topic 1: stipulate earthquake brendan nat brec nds wor spe mitosis quadrant\n  Topic 2: information email please message send mail contact click remove account\n  Topic 3: get us time new offer need free pills money take\n  Topic 4: one use go service market inc business say news international\n  Topic 5: price windows microsoft software professional office xp adobe system cd\n  Topic 6: company statements stock may make look within report investment include\n  Topic 7: font td nbsp height size width align tr color face\n  Topic 8: br right top de oo strong drive body power age\n  Topic 9: com http e www p r b u c computron\n  Topic 10: khumalo injections bf spitfire rn etcmore slutty stereo fre limerick\n\nNumero di Topic: 15\n  Topic 1: tim ects vol ume eff italian mirrior smallcap shaun stoxo\n  Topic 2: email please message send mail contact click remove link line\n  Topic 3: pills order online best save many also today available ship\n  Topic 4: increase power control age mr enter city case open lead\n  Topic 5: price windows microsoft software professional office xp adobe system cd\n  Topic 6: statements time may make look result future number forward could\n  Topic 7: font td nbsp height size width align tr color face\n  Topic 8: br right top de strong drive body st id leave\n  Topic 9: e p r b u c x v n th\n  Topic 10: injections rn duve centertable var valliuum telecom funky zodiac goto\n  Topic 11: get us new one use go offer need money take\n  Topic 12: information stock within report investment securities provide act trade advice\n  Topic 13: company service market include inc business account news share international\n  Topic 14: fw confront dovekie influential ec crania novelty flight wink brendan\n  Topic 15: com http www free computron net info visit html index\n\nNumero di Topic: 20\n  Topic 1: vol ume eff mirrior smallcap shaun stoxo conservative anchorite qui\n  Topic 2: email please message send mail contact click remove link net\n  Topic 3: pills order online best save many also today available ship\n  Topic 4: service market inc business international million high gas state oil\n  Topic 5: price windows microsoft software professional office xp adobe system cd\n  Topic 6: statements may look result future forward could base risk performance\n  Topic 7: font td nbsp height size width align tr color face\n  Topic 8: influential ec flight wel con mon tx radio tournament karpenkov\n  Topic 9: e p r b u c x v n th\n  Topic 10: injections rn duve centertable valliuum telecom funky zodiac goto retiree\n  Topic 11: get new one go offer need money take want work\n  Topic 12: information stock within report investment securities account provide act trade\n  Topic 13: company time make use include number news share pay interest\n  Topic 14: power age mr enter case open lead build property large\n  Topic 15: free computron stop visit show change fax via deal sale\n  Topic 16: com http www border href src nd index php image\n  Topic 17: right oo drive city music die digital head usb load\n  Topic 18: us lot ibm creative canon sony toshiba viewsonic robotics aopen\n  Topic 19: line info br html top de china body st type\n  Topic 20: strong fat female integrate und white dosage black spring class\n\nAnalisi per il modello: model_vec100_win5.model\nCaricamento del modello salvato: model_vec100_win5.model\nVocabolario: com, http, company, price, get, e, www, information, font, email\n\nNumero di Topic: 2\n  Topic 1: font p size b align color face br center style\n  Topic 2: com http company price get e www information email please\n\nNumero di Topic: 3\n  Topic 1: top de strong leave die digital head benefit dr fat\n  Topic 2: company get information email please statements us stock time new\n  Topic 3: com http price e www font td nbsp p height\n\nNumero di Topic: 4\n  Topic 1: de leave die ca female solid integrate un und white\n  Topic 2: get us new one go offer need free service market\n  Topic 3: company information email please statements stock time may make use\n  Topic 4: com http price e www font td nbsp p height\n\nNumero di Topic: 5\n  Topic 1: family top transfer de oo w china strong body power\n  Topic 2: company information statements stock may make look within report service\n  Topic 3: price get e email please us time new one use\n  Topic 4: com http www font td nbsp p height size width\n  Topic 5: smallcap gs promethe dilute grecian pricess greenbriar fm amateur balustrade\n\nNumero di Topic: 6\n  Topic 1: content family top ms de strong body st bill leave\n  Topic 2: one use go service market inc business say number products\n  Topic 3: com http www font td nbsp p height size width\n  Topic 4: price get e us time new offer need free pills\n  Topic 5: smallcap gs grecian abzt accent khumalo dainty draftsman noneedl hs\n  Topic 6: company information email please statements stock may make look message\n\nNumero di Topic: 7\n  Topic 1: p b family top ms de strong body st leave\n  Topic 2: one use go service market inc business say number news\n  Topic 3: com e email please us message free r send u\n  Topic 4: price get time new offer need pills money take c\n  Topic 5: smallcap gs grecian abzt futureevents accent khumalo dainty draftsman noneedl\n  Topic 6: company information statements stock may make look within report investment\n  Topic 7: http www font td nbsp height size width align tr\n\nNumero di Topic: 8\n  Topic 1: top de strong body leave die digital head benefit wireless\n  Topic 2: price x v windows microsoft software professional office xp adobe\n  Topic 3: one use service market inc business say products international million\n  Topic 4: smallcap gs grecian abzt futureevents accent khumalo draftsman noneedl accuse\n  Topic 5: com http e www email please p message r b\n  Topic 6: font td nbsp height size width align tr color face\n  Topic 7: get us time new go offer need free pills money\n  Topic 8: company information statements stock may make look within report investment\n\nNumero di Topic: 9\n  Topic 1: top strong digital benefit dr storage female solid ed integrate\n  Topic 2: price windows microsoft software professional office xp adobe system project\n  Topic 3: one use service market inc business say news products international\n  Topic 4: gs grecian accent khumalo jj tt noneedl hs accuse injections\n  Topic 5: com email please us message free send mail contact click\n  Topic 6: www font td nbsp height width align tr color face\n  Topic 7: get time new go offer need pills money take order\n  Topic 8: company information statements stock may make look within report investment\n  Topic 9: http e p size r b u c x v\n\nNumero di Topic: 10\n  Topic 1: strong die female solid integrate white dosage black port spring\n  Topic 2: price windows microsoft software professional office xp adobe system full\n  Topic 3: first top transfer think china power computer support age mr\n  Topic 4: gs satisfactory khumalo noneedl injections specialist maryland prevention pike cigarettes\n  Topic 5: com email please message send mail contact click remove account\n  Topic 6: www font td nbsp height width align tr color face\n  Topic 7: get us time new one go offer need free pills\n  Topic 8: information statements stock may make look within report investment securities\n  Topic 9: http e p size r b u c x v\n  Topic 10: company use service market include inc business say number news\n\nNumero di Topic: 15\n  Topic 1: nascar motor indianapolis ne silicon flight tie diverse abuse craft\n  Topic 2: windows microsoft software professional office xp adobe special system full\n  Topic 3: say first think china power computer support age mr enter\n  Topic 4: fre anodic valiumxanaxcialis ie graywacke dru duve mexican govenment htmlbody\n  Topic 5: price may news products expect believe project term show next\n  Topic 6: font td nbsp height width align tr color face border\n  Topic 7: get time new one go offer need free pills money\n  Topic 8: company information stock within report investment include securities provide act\n  Topic 9: com http e p r b u c x v\n  Topic 10: statements make look result future number forward could base risk\n  Topic 11: email please us message send mail contact click remove account\n  Topic 12: br top strong body leave die head usb load wireless\n  Topic 13: www computron via cs duty ali lot epson tel zone\n  Topic 14: size family female solid integrate und dosage black spring class\n  Topic 15: use service market inc business share international million internet high\n\nNumero di Topic: 20\n  Topic 1: flight zenith iv salaam bequest allot adventure muller abet flatulent\n  Topic 2: windows software professional office xp adobe system cd ms pro\n  Topic 3: strong digital benefit light dr sport air se fire art\n  Topic 4: fre anodic valiumxanaxcialis graywacke dru duve mexican govenment htmlbody boys\n  Topic 5: price may news products expect believe project term show next\n  Topic 6: font td nbsp height width align tr color face border\n  Topic 7: get new go offer need money take order want work\n  Topic 8: statements look future forward mean fact action events estimate involve\n  Topic 9: http e p r b u c x v n\n  Topic 10: time make include result number without pay could base interest\n  Topic 11: email please message send mail contact click remove link net\n  Topic 12: pills viagra cialis drug soft meds hours health brand mg\n  Topic 13: com www free computron dollars fax via sale duty ali\n  Topic 14: size family female solid integrate und dosage black spring class\n  Topic 15: service market inc business share international million high gas oil\n  Topic 16: microsoft cs lot epson intel ibm creative canon sony toshiba\n  Topic 17: say first think power control age enter city case lead\n  Topic 18: company information stock within report investment securities provide act trade\n  Topic 19: br top body type leave die head usb wireless ca\n  Topic 20: us one use account would thank call give stop state\n\nAnalisi per il modello: model_vec200_win3.model\nCaricamento del modello salvato: model_vec200_win3.model\nVocabolario: com, http, company, price, get, e, www, information, font, email\n\nNumero di Topic: 2\n  Topic 1: proletariat emotional bronco superb cognitive baseboard catherine bramble metabole swamp\n  Topic 2: com http company price get e www information font email\n\nNumero di Topic: 3\n  Topic 1: relaxation madras meson ural hou ects reas rien expe kno\n  Topic 2: com http www font td nbsp p height size width\n  Topic 3: company price get e information email please statements us stock\n\nNumero di Topic: 4\n  Topic 1: increase right top de oo health strong pain drive weight\n  Topic 2: com http e www font td nbsp p height size\n  Topic 3: company price get information email please statements us stock time\n  Topic 4: wn leaden vol asp eff ger mmsr dividend hookup mirrior\n\nNumero di Topic: 5\n  Topic 1: right top de strong pain drive weight power age mr\n  Topic 2: com http e www font td nbsp p height size\n  Topic 3: company information email please statements stock time may make use\n  Topic 4: mmsr smallcap shaun limitation gs nl penls scan elide johnson\n  Topic 5: price get us new one go offer need free service\n\nNumero di Topic: 6\n  Topic 1: top de leave female solid ce und dosage black pi\n  Topic 2: market inc business say international million high gas first system\n  Topic 3: company information statements stock may make use look within report\n  Topic 4: smallcap df abzt marketpiace gram ipod alpine khumalo incumbent raleigh\n  Topic 5: price get e email please us time new one go\n  Topic 6: com http www font td nbsp p height size width\n\nNumero di Topic: 7\n  Topic 1: female und dosage black edit pop silver las debug wilson\n  Topic 2: market inc business million high gas first energy increase well\n  Topic 3: company information statements stock may make use look within report\n  Topic 4: smallcap ipod khumalo raleigh noneedl hs injections dystrophy qi silence\n  Topic 5: get e email please us time new one go offer\n  Topic 6: com http www font td nbsp p height size width\n  Topic 7: price x v windows microsoft software professional office xp adobe\n\nNumero di Topic: 8\n  Topic 1: twit prosaic chalcocite bridegroom acquisitive kinesic emma stipulate earthquake invidious\n  Topic 2: b br right top de strong body power age mr\n  Topic 3: injections blur pike induce rn slutty aeruginosa stereo fre bankrupt\n  Topic 4: e information email please stock message within report send investment\n  Topic 5: com http www font td nbsp height size width computron\n  Topic 6: company statements time may make look result future number forward\n  Topic 7: price us p offer pills r u c order online\n  Topic 8: get new one use go need free service money market\n\nNumero di Topic: 9\n  Topic 1: twit dupe conciliatory prosaic chalcocite bridegroom acquisitive kinesic emma apace\n  Topic 2: th br right top de strong pain body power age\n  Topic 3: injections blur pike induce rn slutty aeruginosa stereo fre bankrupt\n  Topic 4: com http e www email please us message free r\n  Topic 5: font td nbsp height size width b align tr color\n  Topic 6: statements may make look result future number forward could base\n  Topic 7: price p offer pills u c save x v n\n  Topic 8: get time new one use go need service money market\n  Topic 9: company information stock within report investment include securities account news\n\nNumero di Topic: 10\n  Topic 1: acquisitive stipulate earthquake invidious sri spe mitosis quadrant flatulent eagle\n  Topic 2: br right top de oo strong drive body power age\n  Topic 3: injections blur pike induce rn slutty aeruginosa fre bankrupt limerick\n  Topic 4: information email please message send mail contact click remove account\n  Topic 5: font td nbsp height size width b align tr color\n  Topic 6: company statements stock may make look within report investment include\n  Topic 7: price v windows microsoft software professional office xp adobe system\n  Topic 8: get us time new offer need pills money take order\n  Topic 9: one use go service market inc work business say number\n  Topic 10: com http e www p free r u c computron\n\nNumero di Topic: 15\n  Topic 1: asp ger mmsr dividend hookup mirrior wakey smallcap shaun stoxo\n  Topic 2: power computer support age mr enter city case card lead\n  Topic 3: injections induce fre duve mexican boys internetstore thindata eo telecom\n  Topic 4: company make use include account number news share pay interest\n  Topic 5: http www font td nbsp height width align tr color\n  Topic 6: statements may look result future forward could base risk performance\n  Topic 7: pills save viagra cialis drug soft meds hours per paliourg\n  Topic 8: service market inc business say international million high gas first\n  Topic 9: information stock within report investment securities provide act trade advice\n  Topic 10: e p r b u c x v n th\n  Topic 11: price windows microsoft software professional office xp adobe system cd\n  Topic 12: com email please message send mail contact click remove computron\n  Topic 13: get us time new one go offer need free money\n  Topic 14: size info br html right family top de china strong\n  Topic 15: dosage spring alone pop silver conversion vegas triple motor explore\n\nNumero di Topic: 20\n  Topic 1: asp mmsr dividend hookup mirrior wakey smallcap shaun stoxo ladies\n  Topic 2: power computer support age mr enter city case card open\n  Topic 3: injections fre duve mexican boys internetstore eo telecom zdrive zodiac\n  Topic 4: company make use include account news share pay interest expect\n  Topic 5: font td nbsp height size width align tr color face\n  Topic 6: may result future number could base risk performance fact due\n  Topic 7: price us offer need free money order online best save\n  Topic 8: service market inc business international million high gas oil energy\n  Topic 9: information stock report investment securities provide trade advice invest newsletter\n  Topic 10: e p r b c x v n g l\n  Topic 11: windows microsoft software professional office xp adobe system cd ms\n  Topic 12: email please message send mail contact click remove link list\n  Topic 13: get time new one go take want work like say\n  Topic 14: br right top de strong drive body st leave die\n  Topic 15: spring alone pop silver conversion triple explore fw wilson zero\n  Topic 16: net line html content paliourg mobile text federal type format\n  Topic 17: statements look within forward act section mean contain understand material\n  Topic 18: u products th without dollars change canada pass property usa\n  Topic 19: com http www computron info src nd fax index via\n  Topic 20: pills viagra cialis soft hours per health mg hi pain\n\nAnalisi per il modello: model_vec200_win5.model\nCaricamento del modello salvato: model_vec200_win5.model\nVocabolario: com, http, company, price, get, e, www, information, font, email\n\nNumero di Topic: 2\n  Topic 1: font p size b align tr color face href pt\n  Topic 2: com http company price get e www information email please\n\nNumero di Topic: 3\n  Topic 1: de strong city leave die digital head benefit dr co\n  Topic 2: company get information email please statements us stock time new\n  Topic 3: com http price e www font td nbsp p height\n\nNumero di Topic: 4\n  Topic 1: de leave die ca female solid integrate belize un und\n  Topic 2: company information email please statements stock time may make use\n  Topic 3: com http price e www font td nbsp p height\n  Topic 4: get us new one go offer need free market take\n\nNumero di Topic: 5\n  Topic 1: line family top de oo china strong mobile body power\n  Topic 2: company information statements stock may make use look within report\n  Topic 3: com http www font td nbsp p height size width\n  Topic 4: price get e email please us time new one go\n  Topic 5: gs picnic penls beget grecian df greenbriar walter amende bisque\n\nNumero di Topic: 6\n  Topic 1: b line content family top de china strong mobile body\n  Topic 2: company information email please statements stock may make look message\n  Topic 3: com http www font td nbsp p height size width\n  Topic 4: price get e us time new one go offer need\n  Topic 5: gs grecian abzt walter thatexpress futureevents khumalo dainty noneedl column\n  Topic 6: use service market inc business say number news international million\n\nNumero di Topic: 7\n  Topic 1: b family top de w china k strong body age\n  Topic 2: company information statements stock may make look within report investment\n  Topic 3: com e email please message r send u mail c\n  Topic 4: price get us time new go offer need free pills\n  Topic 5: gs grecian abzt thatexpress futureevents khumalo dainty noneedl column colloquy\n  Topic 6: one use service market inc business say number news international\n  Topic 7: http www font td nbsp p height size width align\n\nNumero di Topic: 8\n  Topic 1: b th f top de china strong body st city\n  Topic 2: company information statements stock may make look within report investment\n  Topic 3: service market inc business say news products international million high\n  Topic 4: gs grecian abzt thatexpress khumalo dainty noneedl column colloquy upper\n  Topic 5: com http www font td nbsp p height size width\n  Topic 6: get e email please us time new one use go\n  Topic 7: pills r c order online best save many also n\n  Topic 8: price u x v windows microsoft software professional info office\n\nNumero di Topic: 9\n  Topic 1: top china strong body st city leave die digital head\n  Topic 2: company information statements stock may look within report investment include\n  Topic 3: service market inc business news products international million high gas\n  Topic 4: gs grecian khumalo dainty noneedl column upper injections cute dystrophy\n  Topic 5: font td nbsp height size width align tr color face\n  Topic 6: get email please us time new make one use go\n  Topic 7: pills order online best save many also today available ship\n  Topic 8: e p r b c x v n windows software\n  Topic 9: com http price www free u contact computron link net\n\nNumero di Topic: 10\n  Topic 1: b line th br family de china strong body la\n  Topic 2: company information statements stock may look within report investment include\n  Topic 3: service market inc business say news products international million high\n  Topic 4: gs beget grecian walter amende bisque bedridden claremont debate khumalo\n  Topic 5: font td nbsp height size width align tr color face\n  Topic 6: get e email please us time new make one use\n  Topic 7: pills order online best save many also today available ship\n  Topic 8: p r c x v n windows microsoft software professional\n  Topic 9: com http price www u contact computron net list dollars\n  Topic 10: top st city leave music die digital head benefit dr\n\nNumero di Topic: 15\n  Topic 1: solid integrate und white dosage black spring alone pop silver\n  Topic 2: information statements may look within report investment include securities result\n  Topic 3: company stock service market inc business news provide share international\n  Topic 4: gs khumalo noneedl upper injections maryland bf prevention soil compiiance\n  Topic 5: td height width align tr border href src nd pt\n  Topic 6: get us time new make one use go offer need\n  Topic 7: pills ship viagra cialis drug soft prescription meds hours paliourg\n  Topic 8: font nbsp p size b line color face br center\n  Topic 9: e r u c v n th without g l\n  Topic 10: say first program transfer part power country control computer support\n  Topic 11: http x info f de oo w php k biz\n  Topic 12: com www computron dollars fax via sale duty ali lot\n  Topic 13: email please message send mail contact click remove account link\n  Topic 14: price windows microsoft software professional office xp adobe system cd\n  Topic 15: age city digital benefit dr reduce ed bottom movies jan\n\nNumero di Topic: 20\n  Topic 1: female solid integrate und dosage black spring shell alone pop\n  Topic 2: statements may look within report investment securities result future forward\n  Topic 3: company stock service market include inc business news provide share\n  Topic 4: dru duve mexican htmlbody wl boys internetstore dolars funky goto\n  Topic 5: td height width align tr border href src nd pt\n  Topic 6: get time new make one use go money take want\n  Topic 7: pills viagra cialis soft hours health mg pain weight tabs\n  Topic 8: font nbsp size color face br center style family strong\n  Topic 9: u th without change title notice canada pass property usa\n  Topic 10: program transfer part fund country support age mr enter view\n  Topic 11: die head usb co ca cable storage bite front un\n  Topic 12: com www free computron dollars fax via sale duty ali\n  Topic 13: information email please message send mail contact click remove account\n  Topic 14: http line info html top index php image china biz\n  Topic 15: city digital benefit dr channel ed bottom picture jan kill\n  Topic 16: price us offer need order online best save products many\n  Topic 17: windows microsoft software professional office xp adobe system cd ms\n  Topic 18: e p r b c x v n g l\n  Topic 19: ne flight nat foot zenith anne bequest allot adventure muller\n  Topic 20: high energy increase well year industry network power control computer\n\nTop 10 Cosine Similarity:\n             model_name  n_topics  topic_1  topic_2  cosine_similarity  \\\n271    model_vec50_win3        20        1        3           0.999990   \n166    model_vec50_win3        15        1        3           0.999989   \n1417  model_vec200_win3         7        1        4           0.999986   \n738   model_vec100_win3        20        1       10           0.999970   \n633   model_vec100_win3        15        1       10           0.999962   \n1402  model_vec200_win3         6        1        4           0.999948   \n1651  model_vec200_win3        20        1        3           0.999932   \n483   model_vec100_win3         6        1        5           0.999921   \n1546  model_vec200_win3        15        1        3           0.999910   \n1192  model_vec100_win5        20        1        4           0.999907   \n\n      euclidean_distance  \n271             0.211988  \n166             0.259208  \n1417            0.451314  \n738             0.246665  \n633             0.252614  \n1402            0.484643  \n1651            0.263208  \n483             0.501343  \n1546            0.268195  \n1192            0.261026  \n\nBottom 10 Cosine Similarity:\n             model_name  n_topics  topic_1  topic_2  cosine_similarity  \\\n2148  model_vec200_win5        20        3        5           0.308221   \n1278  model_vec100_win5        20        6       10           0.323022   \n2134  model_vec200_win5        20        2        8           0.323541   \n2033  model_vec200_win5        15        3        5           0.323814   \n1934  model_vec200_win5         9        2        5           0.329272   \n1971  model_vec200_win5        10        2        5           0.334939   \n2021  model_vec200_win5        15        2        5           0.335520   \n2151  model_vec200_win5        20        3        8           0.339583   \n2131  model_vec200_win5        20        2        5           0.343459   \n1309  model_vec100_win5        20        8       16           0.343964   \n\n      euclidean_distance  \n2148            4.483481  \n1278            4.087314  \n2134            3.707233  \n2033            4.350453  \n1934            3.948797  \n1971            4.048055  \n2021            4.572542  \n2151            3.230401  \n2131            4.629296  \n1309            4.488446  \n
\n
\n
\n
\n
" ] } } ], "source": [ "from scipy.spatial.distance import cosine, euclidean\n", "\n", "\"\"\"\n", "Itero sui modelli Word2Vec costruiti in precedenza e su più numeri di topic, applicando KMeans per ogni configurazione.\n", "\"\"\"\n", "\n", "# Numero di topic da testare\n", "n_topics_list = [2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20]\n", "\n", "# Lista per salvare i risultati\n", "results_distances = []\n", "\n", "# Itero sui modelli\n", "for config in configs:\n", " model_name = f\"{config['name']}.model\"\n", " print_colored(f\"\\nAnalisi per il modello: {model_name}\", \"blue\")\n", " \n", " # Carico o addestro il modello Word2Vec\n", " if os.path.exists(model_name):\n", " print_colored(f\"Caricamento del modello salvato: {model_name}\", \"red\")\n", " word2vec_model = Word2Vec.load(model_name)\n", " else:\n", " print_colored(f\"Addestramento del modello: {config['name']}\", \"blue\")\n", " word2vec_model = Word2Vec(\n", " sentences=tokenized_texts,\n", " vector_size=config[\"vector_size\"],\n", " window=config[\"window\"],\n", " min_count=config[\"min_count\"],\n", " workers=multiprocessing.cpu_count(),\n", " sg=config[\"sg\"]\n", " )\n", " word2vec_model.save(model_name)\n", " print_colored(f\"Modello salvato come: {model_name}\", \"red\")\n", " \n", " vocab_sample = list(word2vec_model.wv.index_to_key)[:10]\n", " print_colored(\"Vocabolario: \" + \", \".join(vocab_sample), \"blue\")\n", " \n", " # Itero sui numeri di topic\n", " for n_topics in n_topics_list:\n", " print_colored(f\"\\nNumero di Topic: {n_topics}\", \"blue\")\n", " \n", " # Estrazione dei vettori delle parole\n", " word_vectors = np.array([word2vec_model.wv[word] for word in word2vec_model.wv.index_to_key])\n", " words = word2vec_model.wv.index_to_key\n", " \n", " # Clustering con KMeans\n", " kmeans = KMeans(n_clusters=n_topics, random_state=42)\n", " labels = kmeans.fit_predict(word_vectors)\n", " \n", " # Raggruppo le parole per cluster\n", " topics = []\n", " for cluster in range(n_topics):\n", " words_in_cluster = [words[i] for i, label in enumerate(labels) if label == cluster]\n", " topics.append(words_in_cluster[:10]) # Prime 10 parole del topic\n", " \n", " for i, topic in enumerate(topics):\n", " print(f\" Topic {i + 1}: {' '.join(topic)}\")\n", " \n", " # Calcolo delle distanze semantiche e salvataggio nei risultati\n", " centroids = kmeans.cluster_centers_\n", " for i in range(n_topics):\n", " for j in range(i + 1, n_topics):\n", " cosine_dist = cosine(centroids[i], centroids[j])\n", " euclidean_dist = euclidean(centroids[i], centroids[j])\n", " results_distances.append({\n", " \"model_name\": config['name'],\n", " \"n_topics\": n_topics,\n", " \"topic_1\": i + 1,\n", " \"topic_2\": j + 1,\n", " \"cosine_similarity\": 1 - cosine_dist,\n", " \"euclidean_distance\": euclidean_dist\n", " })\n", "\n", "# Converto i risultati in DataFrame\n", "results_df_distances = pd.DataFrame(results_distances)\n", "\n", "# Filtro i risultati per cosine similarity\n", "top_cosine = results_df_distances.sort_values(by=\"cosine_similarity\", ascending=False).head(10) # Top 10\n", "low_cosine = results_df_distances.sort_values(by=\"cosine_similarity\", ascending=True).head(10) # Bottom 10\n", "\n", "print_colored(\"\\nTop 10 Cosine Similarity:\", \"blue\")\n", "print(top_cosine)\n", "\n", "print_colored(\"\\nBottom 10 Cosine Similarity:\", \"blue\")\n", "print(low_cosine)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Esplorazione e visualizzazione dei dati ottenuti con t-SNE e UMAP

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\n\n\nVisualizzazione per il modello: model_vec50_win3.model\n\nEsecuzione di t-SNE...\n\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\nEsecuzione di UMAP...\n\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\n\n\nVisualizzazione per il modello: model_vec100_win3.model\n\nEsecuzione di t-SNE...\n\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\nEsecuzione di UMAP...\n\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\n\n\nVisualizzazione per il modello: model_vec100_win5.model\n\nEsecuzione di t-SNE...\n\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\nEsecuzione di UMAP...\n\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\n\n\nVisualizzazione per il modello: model_vec200_win3.model\n\nEsecuzione di t-SNE...\n\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\nEsecuzione di UMAP...\n\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\n\n\nVisualizzazione per il modello: model_vec200_win5.model\n\nEsecuzione di t-SNE...\n\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\nEsecuzione di UMAP...\n\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
" ] } } ], "source": [ "from sklearn.manifold import TSNE\n", "from umap.umap_ import UMAP\n", "import random\n", "\n", "# Numero massimo di parole visualizzate\n", "max_words = 150\n", "\n", "# Itero sui modelli\n", "for config in configs:\n", " model_name = f\"{config['name']}.model\"\n", " print_colored(f\"\\n\\n\\nVisualizzazione per il modello: {model_name}\", \"blue\")\n", " \n", " # Carico il modello salvato\n", " if os.path.exists(model_name):\n", " word2vec_model = Word2Vec.load(model_name)\n", " else:\n", " print(f\"Modello non trovato: {model_name}\")\n", " continue \n", " \n", " # Campiono casualmente un sottoinsieme di parole\n", " words = random.sample(word2vec_model.wv.index_to_key, min(max_words, len(word2vec_model.wv.index_to_key)))\n", " word_vectors = np.array([word2vec_model.wv[word] for word in words])\n", "\n", " print(\"\\nEsecuzione di t-SNE...\\n\")\n", " tsne = TSNE(n_components=2, random_state=42)\n", " reduced_vectors_tsne = tsne.fit_transform(word_vectors)\n", "\n", " plt.figure(figsize=(10, 10))\n", " plt.scatter(reduced_vectors_tsne[:, 0], reduced_vectors_tsne[:, 1], alpha=0.6)\n", " for i, word in enumerate(words):\n", " plt.annotate(word, (reduced_vectors_tsne[i, 0], reduced_vectors_tsne[i, 1]), fontsize=8)\n", " plt.title(f\"t-SNE: {model_name}\", fontsize=18)\n", " plt.xticks(fontsize=14, color=\"#b81414\")\n", " plt.yticks(fontsize=14, color=\"#b81414\")\n", " plt.show()\n", " print()\n", "\n", "\n", " print(\"\\nEsecuzione di UMAP...\\n\")\n", " umap_reducer = UMAP(random_state=42)\n", " reduced_vectors_umap = umap_reducer.fit_transform(word_vectors)\n", "\n", " plt.figure(figsize=(10, 10))\n", " plt.scatter(reduced_vectors_umap[:, 0], reduced_vectors_umap[:, 1], alpha=0.6)\n", " for i, word in enumerate(words):\n", " plt.annotate(word, (reduced_vectors_umap[i, 0], reduced_vectors_umap[i, 1]), fontsize=8)\n", " plt.title(f\"UMAP: {model_name}\", fontsize=18)\n", " plt.xticks(fontsize=14, color=\"#b81414\")\n", " plt.yticks(fontsize=14, color=\"#b81414\")\n", " plt.show()\n", " print()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Scelta dei modelli

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

La scelta del modello e del numero di topic dipende dall'obiettivo dell'analisi:


- Individuazione dei Topic Principali -

L'obiettivo è identificare chiaramente i topic principali con elevata coerenza semantica. Il criterio di selezione seguito è stato quello di scegliere il miglior modello basato sul Coherence Score per misurare la coerenza semantica delle parole all'interno dei topic. Un punteggio alto del Coherence Score indica che le parole del topic sono semanticamente correlate e rappresentano un concetto chiaro.

- Calcolo della distanza Semantica -

L'obiettivo è quello di valutare le distanze semantiche tra i topic, utile per capire quanto sono correlati o distinti. Il criterio di selezione seguito è stato quello di scegliere il miglior modello basato sulla metrica di Silhouette la quale indica quanto i cluster sono distanti tra loro. Per questo obiettivo, ci interessa che i cluster abbiano una buona separazione, dato che vogliamo analizzare le distanze semantiche tra i topic.
Il Coherence Score deve essere accettabile per garantire che i topic siano semanticamente significativi ma non è prioritario.

- Analisi Contenutistica Approfondita -

L'obiettivo è quello di estrarre un numero elevato di topic per effettuare un'analisi approfondita dei contenuti. La metrica utilizzata per scegliere il modello è stato il Coherence Score, fondamentale per assicurarsi che i topic siano semanticamente coerenti, dato che l'analisi deve riflettere accuratamente i temi principali. Anche con molti topic occorre un punteggio di coerenza elevato per garantire che ogni gruppo di parole sia significativo.
Il Silhouette Score è meno importante, perché con un numero elevato di topic, i cluster tendono naturalmente a sovrapporsi.

- Valutazione dell'eterogeneità -

L'obiettivo è misurare quanto i topic siano diversi tra loro, cioè quanto siano ben separati semanticamente. La metrica utilizzata è la Silhouette per misurare quanto i cluster (topic) siano ben separati. Un punteggio alto indica che i topic sono distinti e ben definiti. Questo è cruciale per valutare l'eterogeneità, dato che stiamo misurando la separazione tra i topic.

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nIndividuare Topic Principali, model_vec200_win5.model\nTopic 1: font p size b align tr color face href pt\nTopic 2: com http company price get e www information email please\n\nCalcolare Distanza Semantica, model_vec200_win3.model\nDistanza Cosine tra Topic 1 e Topic 2: 0.9802\n\nAnalisi Contenutistica, model_vec200_win5.model\nAnalisi approfondita del Topic 1: solid integrate und white dosage black spring alone pop silver conversion triple nascar motor camera\nAnalisi approfondita del Topic 2: information statements may look within report investment include securities result future forward act could base\nAnalisi approfondita del Topic 3: company stock service market inc business news provide share international million high gas trade expect\nAnalisi approfondita del Topic 4: gs khumalo noneedl upper injections maryland bf prevention soil compiiance spark rn etcmore slutty agaln\nAnalisi approfondita del Topic 5: td height width align tr border href src nd pt style index bgcolor gif cs\nAnalisi approfondita del Topic 6: get us time new make one use go offer need free money take order want\nAnalisi approfondita del Topic 7: pills ship viagra cialis drug soft prescription meds hours paliourg health brand mg hi less\nAnalisi approfondita del Topic 8: font nbsp p size b line color face br center html family top image china\nAnalisi approfondita del Topic 9: e r u c v n th without g l section h change title notice\nAnalisi approfondita del Topic 10: say first program transfer part power country control computer support mr enter view case open\nAnalisi approfondita del Topic 11: http x info f de oo w php k biz st type id die head\nAnalisi approfondita del Topic 12: com www computron dollars fax via sale duty ali lot epson tel zone intel exactly\nAnalisi approfondita del Topic 13: email please message send mail contact click remove account link net list receive address reply\nAnalisi approfondita del Topic 14: price windows microsoft software professional office xp adobe system cd ms pro download drive retail\nAnalisi approfondita del Topic 15: age city digital benefit dr reduce ed bottom movies jan either red kill sun green\n\nValutare Eterogeneità, model_vec200_win3.model\nSilhouette Score (da results_df_main): 0.5074\n
\n
\n
\n
\n
" ] } } ], "source": [ "models_to_use = {\n", " \"Individuare Topic Principali\": {\"model_name\": \"model_vec200_win5\", \"n_topics\": 2},\n", " \"Calcolare Distanza Semantica\": {\"model_name\": \"model_vec200_win3\", \"n_topics\": 2},\n", " \"Analisi Contenutistica\": {\"model_name\": \"model_vec200_win5\", \"n_topics\": 15},\n", " \"Valutare Eterogeneità\": {\"model_name\": \"model_vec200_win3\", \"n_topics\": 5}\n", "}\n", "\n", "# Itero e applico il modello corrispondente\n", "for objective, config in models_to_use.items():\n", " model_name = f\"{config['model_name']}.model\"\n", " print_colored(f\"\\n{objective}, {model_name}\", \"blue\")\n", " \n", " # Carico il modello Word2Vec\n", " word2vec_model = Word2Vec.load(model_name)\n", " \n", " # Estrazione dei vettori\n", " word_vectors = np.array([word2vec_model.wv[word] for word in word2vec_model.wv.index_to_key])\n", " words = word2vec_model.wv.index_to_key\n", " \n", " # Clustering K-Means\n", " n_topics = config[\"n_topics\"]\n", " kmeans = KMeans(n_clusters=n_topics, random_state=42)\n", " labels = kmeans.fit_predict(word_vectors)\n", " \n", " # Analisi specifica\n", " if objective == \"Individuare Topic Principali\":\n", " for cluster in range(n_topics):\n", " words_in_cluster = [words[i] for i, label in enumerate(labels) if label == cluster]\n", " print(f\"Topic {cluster + 1}: {' '.join(words_in_cluster[:10])}\")\n", " \n", " elif objective == \"Calcolare Distanza Semantica\":\n", " centroids = kmeans.cluster_centers_\n", " for i in range(n_topics):\n", " for j in range(i + 1, n_topics):\n", " cosine_sim = 1 - cosine(centroids[i], centroids[j])\n", " print(f\"Distanza Cosine tra Topic {i + 1} e Topic {j + 1}: {cosine_sim:.4f}\")\n", " \n", " elif objective == \"Analisi Contenutistica\":\n", " for cluster in range(n_topics):\n", " words_in_cluster = [words[i] for i, label in enumerate(labels) if label == cluster]\n", " print(f\"Analisi approfondita del Topic {cluster + 1}: {' '.join(words_in_cluster[:15])}\")\n", " \n", " elif objective == \"Valutare Eterogeneità\":\n", " # Recupero il Silhouette Score da results_df\n", " filtered_df = results_df_main[\n", " (results_df_main[\"model_name\"] == config[\"model_name\"]) & (results_df_main[\"n_topics\"] == n_topics)\n", " ]\n", " if filtered_df.empty:\n", " print(f\"⚠️ Nessun dato trovato per il modello: {config['model_name']} con {n_topics} topic.\")\n", " silhouette_avg = None\n", " else:\n", " silhouette_avg = filtered_df[\"silhouette\"].values[0]\n", " print(f\"Silhouette Score (da results_df_main): {silhouette_avg:.4f}\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Per quanto riguarda l'individuazione dei topic principali contenuti dalle email SPAM, l'output è rappresentativo dei contenuti generali di formattazione HTML e di termini associati a comunicazioni aziendali.

Il valore della distanza semantica è molto elevato indicando che esiste una forte similarità semantica. I due topic hanno contenuti sovrapponibili provenienti da contesti simili.

Per quanto riguarda l'analisi contenuntistica approfondita, il modello scelto può essere sfruttato per identificare categorie specifiche di contenuti spam.

L'eterogeneità dei contenuti è moderata indicando che i cluster non sono completamente separati.

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Word Cloud

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\n\n\n\nGenerazione Word Cloud per il modello: model_vec200_win5.model\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\n\n\nGenerazione Word Cloud per il modello: model_vec200_win3.model\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\n\n\nGenerazione Word Cloud per il modello: model_vec200_win5.model\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\n\n\nGenerazione Word Cloud per il modello: model_vec200_win3.model\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "from wordcloud import WordCloud\n", "import matplotlib.pyplot as plt\n", "\n", "# Seleziono i modelli scelti\n", "selected_models = [\n", " {\"model_name\": \"model_vec200_win5.model\", \"n_topics\": 2, \"description\": \"Topic Principali\"},\n", " {\"model_name\": \"model_vec200_win3.model\", \"n_topics\": 2, \"description\": \"Calcolo della Distanza Semantica\"},\n", " {\"model_name\": \"model_vec200_win5.model\", \"n_topics\": 15, \"description\": \"Analisi Contenutistica\"},\n", " {\"model_name\": \"model_vec200_win3.model\", \"n_topics\": 5, \"description\": \"Valutazione dell’Eterogeneità\"}, \n", "]\n", "\n", "# Itero sui modelli selezionati\n", "for model_config in selected_models:\n", " model_name = model_config[\"model_name\"]\n", " n_topics = model_config[\"n_topics\"]\n", " description = model_config[\"description\"]\n", " \n", " print_colored(f\"\\n\\n\\n\\nGenerazione Word Cloud per il modello: {model_name}\", \"blue\")\n", " \n", " # Carico il modello salvato\n", " word2vec_model = Word2Vec.load(model_name)\n", " \n", " # Estraggo i topic con KMeans\n", " word_vectors = np.array([word2vec_model.wv[word] for word in word2vec_model.wv.index_to_key])\n", " words = word2vec_model.wv.index_to_key\n", " \n", " kmeans = KMeans(n_clusters=n_topics, random_state=42)\n", " labels = kmeans.fit_predict(word_vectors)\n", " \n", " # Imposto la griglia per le nuvole di parole\n", " cols = 2 # Numero di colonne\n", " rows = (n_topics + 1) // cols # Numero di righe\n", " fig, axes = plt.subplots(rows, cols, figsize=(12, rows * 4))\n", " axes = axes.flatten()\n", "\n", " # Genero una Word Cloud per ciascun topic\n", " for cluster in range(n_topics):\n", " words_in_cluster = [words[i] for i, label in enumerate(labels) if label == cluster]\n", " topic_words = \" \".join(words_in_cluster[:50]) # Uso solo le prime 50 parole\n", " \n", " # Genero la Word Cloud\n", " wordcloud = WordCloud(\n", " width=800, height=400, background_color=\"white\", max_words=50, colormap=\"viridis\"\n", " ).generate(topic_words)\n", " \n", " # Subplot\n", " ax = axes[cluster]\n", " ax.imshow(wordcloud, interpolation=\"bilinear\")\n", " ax.axis(\"off\")\n", " ax.set_title(f\"Topic {cluster + 1}\", fontsize=14, color=\"#b81414\")\n", " \n", " # Nascondo subplot vuoti\n", " for ax in axes[n_topics:]:\n", " ax.axis(\"off\")\n", " \n", " fig.suptitle(f\"{description}\\n\", fontsize=22)\n", " plt.tight_layout()\n", " plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Rappresentazione della Distanza Semantica

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

t-SNE è uno strumento utile per rappresentare visivamente le distanze semantiche, per esplorare relazioni tra parole, concetti e cluster.
Può ridurre a 2 o 3 dimensioni i vettori ad alta dimensionalità dei modelli Word2Vec.
Funziona bene per dataset piccoli

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\n\nGenerazione t-SNE per il modello: model_vec200_win5.model (Topic Principali)\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\nGenerazione t-SNE per il modello: model_vec200_win3.model (Calcolo della Distanza Semantica)\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\nGenerazione t-SNE per il modello: model_vec200_win5.model (Analisi Contenutistica)\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\n\nGenerazione t-SNE per il modello: model_vec200_win3.model (Valutazione dell’Eterogeneità)\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "from sklearn.manifold import TSNE\n", "\n", "selected_models = [\n", " {\"model_name\": \"model_vec200_win5.model\", \"n_topics\": 2, \"description\": \"Topic Principali\"},\n", " {\"model_name\": \"model_vec200_win3.model\", \"n_topics\": 2, \"description\": \"Calcolo della Distanza Semantica\"},\n", " {\"model_name\": \"model_vec200_win5.model\", \"n_topics\": 15, \"description\": \"Analisi Contenutistica\"},\n", " {\"model_name\": \"model_vec200_win3.model\", \"n_topics\": 5, \"description\": \"Valutazione dell’Eterogeneità\"}, \n", "]\n", "\n", "# Visualizzazione con t-SNE per ogni modello selezionato\n", "for model_config in selected_models:\n", " model_name = model_config[\"model_name\"]\n", " n_topics = model_config[\"n_topics\"]\n", " description = model_config[\"description\"]\n", " \n", " print_colored(f\"\\n\\nGenerazione t-SNE per il modello: {model_name} ({description})\", \"blue\")\n", " \n", " # Carico il modello Word2Vec\n", " word2vec_model = Word2Vec.load(model_name)\n", " word_vectors = np.array([word2vec_model.wv[word] for word in word2vec_model.wv.index_to_key])\n", " words = word2vec_model.wv.index_to_key\n", "\n", " # Clustering con KMeans\n", " kmeans = KMeans(n_clusters=n_topics, random_state=42)\n", " labels = kmeans.fit_predict(word_vectors)\n", " \n", " # Riduzione dimensionale con t-SNE\n", " tsne = TSNE(n_components=2, random_state=42)\n", " reduced_vectors = tsne.fit_transform(word_vectors)\n", "\n", " plt.figure(figsize=(10, 8))\n", " for i in range(n_topics):\n", " cluster_points = reduced_vectors[labels == i]\n", " plt.scatter(cluster_points[:, 0], cluster_points[:, 1], label=f\"Topic {i + 1}\")\n", " plt.title(f\"\\nVisualizzazione dei Topic con t-SNE ({description})\", fontsize=18)\n", " plt.xticks(fontsize=14, color=\"#b81414\")\n", " plt.yticks(fontsize=14, color=\"#b81414\")\n", " plt.legend()\n", " plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Radar plot

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "from math import pi\n", "\n", "metrics = {\n", " \"Coherence\": [0.756923, 0.7639, 0.5869, 0.6793],\n", " \"Silhouette\": [0.621298, 0.6031, 0.5074, 0.2783],\n", " \"Topic\": [2, 2, 15, 5],\n", " \"Model\": [\"model_vec200_win3\", \"model_vec200_win5\", \"model_vec200_win3_dup\", \"model_vec200_win5_dup\"],\n", "}\n", "labels = metrics[\"Model\"] # Etichette dei modelli\n", "df_metrics = pd.DataFrame(metrics).set_index(\"Model\") # Imposto \"Model\" come indice\n", "\n", "# Radar plot\n", "angles = np.linspace(0, 2 * np.pi, len(df_metrics.columns), endpoint=False).tolist()\n", "angles += angles[:1]\n", "\n", "fig, ax = plt.subplots(figsize=(6, 6), subplot_kw=dict(polar=True))\n", "\n", "# Iterazione sulle righe del DataFrame\n", "for idx, row in df_metrics.iterrows():\n", " values = row.tolist() + row.tolist()[:1] # Chiudo il radar plot\n", " ax.plot(angles, values, label=idx)\n", " ax.fill(angles, values, alpha=0.25)\n", "\n", "ax.set_yticks([])\n", "ax.set_xticks(angles[:-1])\n", "ax.set_xticklabels(df_metrics.columns)\n", "plt.title(\"Radar Plot dei Modelli\", fontsize=16, pad=20)\n", "ax.legend(bbox_to_anchor=(1.2, 1.05))\n", "plt.tight_layout()\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Email NON SPAM

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Estrazione dalle email NON SPAM di informazioni sulle Organizzazioni menzionate

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Dataset iniziale droppato\n                                                text  label_num\n0  Subject: enron methanol ; meter # : 988291\\nth...          0\n1  Subject: hpl nom for january 9 , 2001\\n( see a...          0\n2  Subject: neon retreat\\nho ho ho , we ' re arou...          0\n3  Subject: photoshop , windows , office . cheap ...          1\n4  Subject: re : indian springs\\nthis deal is to ...          0\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Dataset iniziale droppato\",\"blue\")\n", "print(df_email_dropped.head())\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Dataset droppato contenente solo email NON SPAM\n                                                text  label_num\n0  Subject: enron methanol ; meter # : 988291\\nth...          0\n1  Subject: hpl nom for january 9 , 2001\\n( see a...          0\n2  Subject: neon retreat\\nho ho ho , we ' re arou...          0\n4  Subject: re : indian springs\\nthis deal is to ...          0\n5  Subject: ehronline web address change\\nthis me...          0\n
\n
\n
\n
\n
" ] } } ], "source": [ "df_email_dropped_ham = df_email_dropped[df_email_dropped['label_num'] == 0]\n", "\n", "print_colored(\"Dataset droppato contenente solo email NON SPAM\", \"blue\")\n", "print(df_email_dropped_ham.head())\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Analizzo il contenuto delle prime 10 email per comprendere il corpus e il contesto, al fine di definire il flusso di lavoro successivo

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Email 1\nSubject: enron methanol ; meter # : 988291\nthis is a follow up to the note i gave you on monday , 4 / 3 / 00 { preliminary\nflow data provided by daren } .\nplease override pop ' s daily volume { presently zero } to reflect daily\nactivity you can obtain from gas control .\nthis change is needed asap for economics purposes .\n\n\nEmail 2\nSubject: hpl nom for january 9 , 2001\n( see attached file : hplnol 09 . xls )\n- hplnol 09 . xls\n\n\nEmail 3\nSubject: neon retreat\nho ho ho , we ' re around to that most wonderful time of the year - - - neon leaders retreat time !\ni know that this time of year is extremely hectic , and that it ' s tough to think about anything past the holidays , but life does go on past the week of december 25 through january 1 , and that ' s what i ' d like you to think about for a minute .\non the calender that i handed out at the beginning of the fall semester , the retreat was scheduled for the weekend of january 5 - 6 . but because of a youth ministers conference that brad and dustin are connected with that week , we ' re going to change the date to the following weekend , january 12 - 13 . now comes the part you need to think about .\ni think we all agree that it ' s important for us to get together and have some time to recharge our batteries before we get to far into the spring semester , but it can be a lot of trouble and difficult for us to get away without kids , etc . so , brad came up with a potential alternative for how we can get together on that weekend , and then you can let me know which you prefer .\nthe first option would be to have a retreat similar to what we ' ve done the past several years . this year we could go to the heartland country inn ( www . . com ) outside of brenham . it ' s a nice place , where we ' d have a 13 - bedroom and a 5 - bedroom house side by side . it ' s in the country , real relaxing , but also close to brenham and only about one hour and 15 minutes from here . we can golf , shop in the antique and craft stores in brenham , eat dinner together at the ranch , and spend time with each other . we ' d meet on saturday , and then return on sunday morning , just like what we ' ve done in the past .\nthe second option would be to stay here in houston , have dinner together at a nice restaurant , and then have dessert and a time for visiting and recharging at one of our homes on that saturday evening . this might be easier , but the trade off would be that we wouldn ' t have as much time together . i ' ll let you decide .\nemail me back with what would be your preference , and of course if you ' re available on that weekend . the democratic process will prevail - - majority vote will rule ! let me hear from you as soon as possible , preferably by the end of the weekend . and if the vote doesn ' t go your way , no complaining allowed ( like i tend to do ! )\nhave a great weekend , great golf , great fishing , great shopping , or whatever makes you happy !\nbobby\n\n\nEmail 4\nSubject: re : indian springs\nthis deal is to book the teco pvr revenue . it is my understanding that teco\njust sends us a check , i haven ' t received an answer as to whether there is a\npredermined price associated with this deal or if teco just lets us know what\nwe are giving . i can continue to chase this deal down if you need .\n\n\nEmail 5\nSubject: ehronline web address change\nthis message is intended for ehronline users only .\ndue to a recent change to ehronline , the url ( aka \" web address \" ) for accessing ehronline needs to be changed on your computer . the change involves adding the letter \" s \" to the \" http \" reference in the url . the url for accessing ehronline should be : https : / / ehronline . enron . com .\nthis change should be made by those who have added the url as a favorite on the browser .\n\n\nEmail 6\nSubject: spring savings certificate - take 30 % off\nsave 30 % when you use our customer appreciation spring savings\ncertificate at foot locker , lady foot locker , kids foot locker and at\nour online stores !\nwelcome to our customer appreciation spring savings certificate !\nuse the special certificate below and receive 30 % off your purchases either in our stores or online . hurry ! this 4 - day sale begins thursday , march 22 and ends sunday , march 25 .\nshare the savings today and e - mail this offer to your friends . many items already are reduced and the 30 % discount is taken off the lowest sale price .\nclick below to print your customer appreciation spring savings certificate . you must present this coupon at any foot locker , lady foot locker or kids foot locker store in the u . s . foot locker canada is not participating in this program .\nready , set , save !\nour spring savings discount will automatically appear when you use the links below or type camlem 21 into the promotion code box during checkout .\nfootlocker . com certificate code : camlem 21\nladyfootlocker . com certificate code : camlem 21\nkidsfootlocker . com certificate code : camlem 21\nremember , returns are hassle - free . simply bring your items to any of our stores nationwide or through the mail .\ndon ' t be left out - register today to learn about our new products , promotions , events and other specials . simply click below .\nterms and conditions . some exclusions apply , please see manager for complete details . certificate must be presented at the time of purchase and cannot be used in conjunction with any other discount offer or associate benefit . not redeemable for cash . applicable taxes must be paid by bearer . cannot be applied to prior purchases or to gift card purchases . void where prohibited , licensed or regulated . catalog exclusions apply . valid thursday , 3 / 22 / 01 through sunday , 3 / 25 / 01 . foot locker canada will not participate in this program .\nif you do not wish to receive future emails please click below to\nunsubscribe :\n\n\n\nEmail 7\nSubject: noms / actual flow for 2 / 26\nwe agree\n- - - - - - - - - - - - - - - - - - - - - - forwarded by melissa jones / texas utilities on\n02 / 27 / 2001\n10 : 33 am - - - - - - - - - - - - - - - - - - - - - - - - - - -\n\" eileen ponton \" on 02 / 27 / 2001 09 : 46 : 26 am\nto : david avila / lsp / enserch / us @ tu , charlie stone / texas utilities @ tu , melissa\njones / texas utilities @ tu , hpl . scheduling @ enron . com ,\nliz . bellamy @ enron . com\ncc :\nsubject : noms / actual flow for 2 / 26\ndate nom flow - mcf flow - mmbtu\n2 / 26 / 01 0 456 469\nbtu = 1 . 027\n\n\nEmail 8\nSubject: nominations for oct . 21 - 23 , 2000\n( see attached file : hplnl 021 . xls )\n- hplnl 021 . xls\n\n\nEmail 9\nSubject: enron / hpl actuals for august 28 , 2000\nteco tap 20 . 000 / enron ; 120 . 000 / hpl gas daily\nls hpl lsk ic 20 . 000 / enron\n\n\nEmail 10\nSubject: tenaska iv july\ndarren :\nplease remove the price on the tenaska iv sale , deal 384258 , for july and enter the demand fee . the amount should be $ 3 , 902 , 687 . 50 .\nthanks ,\nmegan\n\n\n
\n
\n
\n
\n
" ] } } ], "source": [ "for idx, email in enumerate(df_email_dropped_ham['text'].head(10), 1):\n", " print_colored(f\"Email {idx}\", \"blue\")\n", " print(email)\n", " print(\"\\n\") \n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Prima prova di estrazione delle entità dalle email con tutte le etichette riconosciute da SpaCy per avere un'idea più generale delle informazioni riguardanti le Organizzazioni utilizzando en_core_web_sm

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Email 1 - Entità rilevanti:\n{'enron methanol, 988291, monday, daily, zero, daily'}\nEmail 2 - Entità rilevanti:\n{'january 9 , 2001, 09, xls, 09'}\nEmail 3 - Entità rilevanti:\n{'ho ho ho, january 1, a minute, the weekend of january 5 - 6, that week, the following weekend, january 12 - 13, the spring semester, that weekend, first, the past several years, this year, 13, 5, only about one hour and 15 minutes, saturday, sunday morning, second, houston, saturday, evening, that weekend, the end of the weekend, doesn, a great weekend, bobby'}\nEmail 4 - Entità rilevanti:\n{'indian'}\nEmail 5 - Entità rilevanti:\n{'Nessuna'}\nEmail 6 - Entità rilevanti:\n{'spring, 30 %, 30 %, spring, kids foot locker, spring, 30 %, 4 - day, thursday, march 22, sunday, march 25, today, the 30 %, spring, the u ., foot locker canada, spring, 21, 21, 21, 21, don, today, bearer, thursday, 3 / 22 / 01, foot locker canada'}\nEmail 7 - Entità rilevanti:\n{'2 / 26, melissa jones / texas, 02 / 27 / 2001, 10, 33, 02 / 27 / 2001 09, 46, 26, david avila, charlie, melissa, texas, 2 / 26, 2 / 26 / 01, 1, 027'}\nEmail 8 - Entità rilevanti:\n{'oct . 21 - 23 , 2000, 021, xls, 021'}\nEmail 9 - Entità rilevanti:\n{'august 28 , 2000, 20, 000, 120, daily, 20, 000'}\nEmail 10 - Entità rilevanti:\n{'tenaska, july, tenaska, july, 3, 902, 687, 50'}\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Carico il modello NER pre-addestrato\n", "nlp = spacy.load(\"en_core_web_sm\")\n", "\n", "# Filtro le email NON SPAM\n", "df_email_dropped_ham = df_email_dropped[df_email_dropped['label_num'] == 0]['text']\n", "\n", "# Lista per salvare le entità rilevanti\n", "all_entities = []\n", "\n", "for email in df_email_dropped_ham:\n", " doc = nlp(email) # Applico NER a ciascuna email\n", " relevant_entities = [\n", " ent.text for ent in doc.ents if ent.label_ in {\"ORG\", \"PERSON\", \"GPE\", \"NORP\", \"FAC\",\"LOC\", \"PRODUCT\", \"EVENT\", \"WORK_OF_ART\", \"LAW\",\n", " \"LANGUAGE\", \"DATE\", \"TIME\", \"PERCENT\", \"MONEY\",\"QUANTITY\", \"ORDINAL\", \"CARDINAL\", \"WORK_OF_ART\", \"LAW\"\n", " }\n", " ]\n", " all_entities.append(relevant_entities)\n", "\n", "for idx, entities in enumerate(all_entities[:10]): \n", " print_colored(f\"Email {idx + 1} - Entità rilevanti:\", \"blue\")\n", " print({', '.join(entities) if entities else 'Nessuna'})\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Seconda prova di estrazione utilizzando solo l'etichetta ORG per avere un'idea più precisa sulle Organizzazioni utilizzando en_core_web_sm

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Email 1 - Organizzazioni menzionate: Nessuna\nEmail 2 - Organizzazioni menzionate: xls\nEmail 3 - Organizzazioni menzionate: Nessuna\nEmail 4 - Organizzazioni menzionate: Nessuna\nEmail 5 - Organizzazioni menzionate: Nessuna\nEmail 6 - Organizzazioni menzionate: bearer, foot locker canada\nEmail 7 - Organizzazioni menzionate: Nessuna\nEmail 8 - Organizzazioni menzionate: xls\nEmail 9 - Organizzazioni menzionate: Nessuna\nEmail 10 - Organizzazioni menzionate: Nessuna\nEmail 11 - Organizzazioni menzionate: christy sweeney / hou, darron c\ngiron / hou, riley / hou /, mary m smith / hou /, zivley / hou, austin / hou /, nathan l hlavaty / hou\nEmail 12 - Organizzazioni menzionate: riley / hou /, mike morris / corp / enron @\nEmail 13 - Organizzazioni menzionate: boas / hou\nEmail 14 - Organizzazioni menzionate: Nessuna\nEmail 15 - Organizzazioni menzionate: Nessuna\nEmail 16 - Organizzazioni menzionate: Nessuna\nEmail 17 - Organizzazioni menzionate: xls\nEmail 18 - Organizzazioni menzionate: Nessuna\nEmail 19 - Organizzazioni menzionate: Nessuna\nEmail 20 - Organizzazioni menzionate: Nessuna\nEmail 21 - Organizzazioni menzionate: Nessuna\nEmail 22 - Organizzazioni menzionate: Nessuna\nEmail 23 - Organizzazioni menzionate: neuweiler / hou\nEmail 24 - Organizzazioni menzionate: the houston expl, megan parker / corp / enron\nEmail 25 - Organizzazioni menzionate: Nessuna\nEmail 26 - Organizzazioni menzionate: cst, the enron corp .\nEmail 27 - Organizzazioni menzionate: lozano / hou, north america corp ., the texas gas group\nEmail 28 - Organizzazioni menzionate: Nessuna\nEmail 29 - Organizzazioni menzionate: Nessuna\nEmail 30 - Organizzazioni menzionate: Nessuna\nEmail 31 - Organizzazioni menzionate: ami chokshi / corp / enron @ enron\n, ami chokshi / corp / enron\nEmail 32 - Organizzazioni menzionate: Nessuna\nEmail 33 - Organizzazioni menzionate: doc\nEmail 34 - Organizzazioni menzionate: Nessuna\nEmail 35 - Organizzazioni menzionate: zivley / hou, cico oil & gas\nEmail 36 - Organizzazioni menzionate: Nessuna\nEmail 37 - Organizzazioni menzionate: devers @ columbia, jen black, nyu, michelle, abercrombie, trisha, brenda, sutton, daniel, syscom - inc ., canedy, janelle mccan\nEmail 38 - Organizzazioni menzionate: Nessuna\nEmail 39 - Organizzazioni menzionate: Nessuna\nEmail 40 - Organizzazioni menzionate: Nessuna\nEmail 41 - Organizzazioni menzionate: pg & e, pg & e texas, pg & e texas bulletin board, pg &, sherlyn\nEmail 42 - Organizzazioni menzionate: Nessuna\nEmail 43 - Organizzazioni menzionate: lone star\nEmail 44 - Organizzazioni menzionate: Nessuna\nEmail 45 - Organizzazioni menzionate: Nessuna\nEmail 46 - Organizzazioni menzionate: Nessuna\nEmail 47 - Organizzazioni menzionate: wagner brown\nEmail 48 - Organizzazioni menzionate: aquila\nEmail 49 - Organizzazioni menzionate: Nessuna\nEmail 50 - Organizzazioni menzionate: chevron\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Carico il modello NER\n", "nlp = spacy.load(\"en_core_web_sm\")\n", "\n", "# Filtro le email NON SPAM\n", "df_email_dropped_ham = df_email_dropped[df_email_dropped['label_num'] == 0]['text']\n", "\n", "# Lista per salvare le entità\n", "all_organizations = []\n", "\n", "for email in df_email_dropped_ham:\n", " doc = nlp(email) # Applico NER\n", " # Filtro solo \"ORG\" e rimuovo entità con numeri\n", " organizations = [\n", " ent.text for ent in doc.ents if ent.label_ == \"ORG\" and not any(char.isdigit() for char in ent.text)\n", " ]\n", " all_organizations.append(set(organizations)) # Evito duplicati all'interno della stessa email\n", "\n", "for idx, orgs in enumerate(all_organizations[:50]): # Mostro solo le prime 50 email\n", " print(f\"Email {idx + 1} - Organizzazioni menzionate: {', '.join(orgs) if orgs else 'Nessuna'}\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Terza prova di estrazione utilizzando solo l'etichetta ORG ma questa volta utilizzando en_core_web_trf

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
|===========================================================================|\n|                  PyTorch CUDA memory summary, device ID 0                 |\n|---------------------------------------------------------------------------|\n|            CUDA OOMs: 0            |        cudaMalloc retries: 0         |\n|===========================================================================|\n|        Metric         | Cur Usage  | Peak Usage | Tot Alloc  | Tot Freed  |\n|---------------------------------------------------------------------------|\n| Allocated memory      | 245257 KiB |    948 MiB |   1188 MiB |    948 MiB |\n|---------------------------------------------------------------------------|\n| Active memory         | 245257 KiB |    948 MiB |   1188 MiB |    948 MiB |\n|---------------------------------------------------------------------------|\n| Requested memory      | 243802 KiB |    946 MiB |   1184 MiB |    946 MiB |\n|---------------------------------------------------------------------------|\n| GPU reserved memory   | 462848 KiB |   1040 MiB |   1040 MiB | 602112 KiB |\n|---------------------------------------------------------------------------|\n| Non-releasable memory | 217590 KiB | 231012 KiB |   1135 MiB |    923 MiB |\n|---------------------------------------------------------------------------|\n| Allocations           |      79    |     300    |     379    |     300    |\n|---------------------------------------------------------------------------|\n| Active allocs         |      79    |     300    |     379    |     300    |\n|---------------------------------------------------------------------------|\n| GPU reserved segments |      27    |      50    |      50    |      23    |\n|---------------------------------------------------------------------------|\n| Non-releasable allocs |      27    |      39    |      99    |      72    |\n|---------------------------------------------------------------------------|\n| Oversize allocations  |       0    |       0    |       0    |       0    |\n|---------------------------------------------------------------------------|\n| Oversize GPU segments |       0    |       0    |       0    |       0    |\n|===========================================================================|\n\nGPU utilizzata con successo.\n\n\nEmail 1 - Organizzazioni estratte:\n{'pop'}\n\nEmail 2 - Organizzazioni estratte:\n{'Nessuna'}\n\nEmail 3 - Organizzazioni estratte:\n{'Nessuna'}\n\nEmail 4 - Organizzazioni estratte:\n{'teco'}\n\nEmail 5 - Organizzazioni estratte:\n{'Nessuna'}\n\nEmail 6 - Organizzazioni estratte:\n{'kidsfootlocker . com, footlocker . com, foot locker canada'}\n\nEmail 7 - Organizzazioni estratte:\n{'enron, enserch, texas utilities, tu'}\n\nEmail 8 - Organizzazioni estratte:\n{'Nessuna'}\n\nEmail 9 - Organizzazioni estratte:\n{'hpl gas daily'}\n\nEmail 10 - Organizzazioni estratte:\n{'tenaska iv'}\n\nEmail 11 - Organizzazioni estratte:\n{'sitara, svcs, enron, ees'}\n\nEmail 12 - Organizzazioni estratte:\n{'swift, enronxgate'}\n\nEmail 13 - Organizzazioni estratte:\n{'volume mgmt, none\\n'}\n\nEmail 14 - Organizzazioni estratte:\n{'eops'}\n\nEmail 15 - Organizzazioni estratte:\n{'enron, cc'}\n\nEmail 16 - Organizzazioni estratte:\n{'concorde churchill, equistar channelview'}\n\nEmail 17 - Organizzazioni estratte:\n{'Nessuna'}\n\nEmail 18 - Organizzazioni estratte:\n{'Nessuna'}\n\nEmail 19 - Organizzazioni estratte:\n{'enron, cc, texas utilities, hpl, tu, teco, enron . com, iferc'}\n\nEmail 20 - Organizzazioni estratte:\n{'spinner . com, spinner'}\n\nEmail 21 - Organizzazioni estratte:\n{'hpl, aep'}\n\nEmail 22 - Organizzazioni estratte:\n{'enron, logitech, houston, european resolution center, north american resolution center'}\n\nEmail 23 - Organizzazioni estratte:\n{'sitara, valero'}\n\nEmail 24 - Organizzazioni estratte:\n{'the houston exploration, the houston exploration company, danny'}\n\nEmail 25 - Organizzazioni estratte:\n{'iferc'}\n\nEmail 26 - Organizzazioni estratte:\n{'vanguard lifestrategy, the enron corp, cst, enron'}\n\nEmail 27 - Organizzazioni estratte:\n{'enron north america corp, cc, the texas gas group, ect, eol'}\n\nEmail 28 - Organizzazioni estratte:\n{'Nessuna'}\n\nEmail 29 - Organizzazioni estratte:\n{'Nessuna'}\n\nEmail 30 - Organizzazioni estratte:\n{'Nessuna'}\n\nEmail 31 - Organizzazioni estratte:\n{'cc, copanos'}\n\nEmail 32 - Organizzazioni estratte:\n{'kcs energy, hpl\\n, cc, hpl, bob withers, lst rev'}\n\nEmail 33 - Organizzazioni estratte:\n{'fuel supply, archer\\n'}\n\nEmail 34 - Organizzazioni estratte:\n{'iferc'}\n\nEmail 35 - Organizzazioni estratte:\n{'vance, cico oil & gas co'}\n\nEmail 36 - Organizzazioni estratte:\n{'sitara, cc, samson lone star, svcs, hesco gathering oil co, winn exploration co, hesco, period\\n'}\n\nEmail 37 - Organizzazioni estratte:\n{'strategicweather, aol . com, launidadlatina, syscom - inc ., related . com, newyorklife . com, erac, stevens - tech, bigfoot, jpmorgan . com, stern . nyu, stern, delinvest, excite, summitbank, aol ., huntoon . com, usa, philamuseum, get, yahoo . com, gap, gap store, advanstar . com'}\n\nEmail 38 - Organizzazioni estratte:\n{'Nessuna'}\n\nEmail 39 - Organizzazioni estratte:\n{'Nessuna'}\n\nEmail 40 - Organizzazioni estratte:\n{'enron, the enrononline product control group, enrononline'}\n\nEmail 41 - Organizzazioni estratte:\n{'pg & e, legal, deal, hpl, desk, ena, pg & e texas, pg'}\n\nEmail 42 - Organizzazioni estratte:\n{'Nessuna'}\n\nEmail 43 - Organizzazioni estratte:\n{\"tenaska iv ', lone star, cornhusker - lone star payments\"}\n\nEmail 44 - Organizzazioni estratte:\n{'hpl gas daily'}\n\nEmail 45 - Organizzazioni estratte:\n{'enron, enserch, texas utilities, tu'}\n\nEmail 46 - Organizzazioni estratte:\n{'sitara, just, equistar'}\n\nEmail 47 - Organizzazioni estratte:\n{'enron, tufco, txu'}\n\nEmail 48 - Organizzazioni estratte:\n{'aquila dallas marketing, epgt, unify, it and business group, aquila marketing, el paso'}\n\nEmail 49 - Organizzazioni estratte:\n{'ehronline'}\n\nEmail 50 - Organizzazioni estratte:\n{'sitara, cc, gas settlements, houston pipe line co, hsc _ flw, enronxgate, ect, chevron phillips chemical co, enron . com, hplc, dfarmer'}\n
\n
\n
\n
\n
" ] } } ], "source": [ "torch.cuda.empty_cache() # Libera la memoria inutilizzata\n", "\n", "torch.cuda.set_per_process_memory_fraction(0.8, device=0) \n", "\n", "# Configurazione per l'allocazione della memoria\n", "os.environ[\"PYTORCH_CUDA_ALLOC_CONF\"] = (\n", " \"expandable_segments,\" # Permette all'allocatore di utilizzare segmenti di memoria espandibili per \n", " # ridurre la frammentazione e migliorare la gestione della memoria\n", " \"garbage_collection_threshold:0.75,\" #Specifica la soglia di utilizzo della memoria per \n", " # avviare la raccolta dei frammenti non utilizzati\n", " \"release_cuda_memory:True,\" # Permette di rilasciare memoria GPU inutilizzata automaticamente durante il runtime\n", " \"device_allocator_retry:True\" # Consente di gestire errori di allocazione riprovando una nuova \n", " # allocazione dopo aver svuotato la memoria non utilizzata\n", ")\n", "\n", "cuda_memory_summary = torch.cuda.memory_summary(device=None, abbreviated=True)\n", "print_colored(cuda_memory_summary, \"\")\n", "\n", "# Configuro SpaCy per usare la GPU\n", "set_gpu_allocator(\"pytorch\") # Usa PyTorch come backend per l'allocazione\n", "require_gpu() # Forzo l'uso della GPU\n", "\n", "nlp = spacy.load(\"en_core_web_trf\")\n", "\n", "# Verifico se la GPU è in uso\n", "gpu_status = spacy.prefer_gpu()\n", "if gpu_status:\n", " print_colored(\"GPU utilizzata con successo.\\n\", \"blue\")\n", "else:\n", " print_colored(\"SpaCy sta utilizzando la CPU.\", \"blue\")\n", "\n", "# Filtro le email NON SPAM\n", "df_email_dropped_ham = df_email_dropped[df_email_dropped['label_num'] == 0]['text']\n", "\n", "# Lista per salvare le entità\n", "all_organizations = []\n", "\n", "for email in df_email_dropped_ham:\n", " doc = nlp(email) # Applica NER\n", " # Filtro solo \"ORG\" e rimuovo entità con numeri\n", " organizations = [\n", " ent.text for ent in doc.ents if ent.label_ == \"ORG\" and not any(char.isdigit() for char in ent.text)\n", " ]\n", " all_organizations.append(set(organizations)) # Evito duplicati all'interno della stessa email\n", "\n", "for idx, orgs in enumerate(all_organizations[:50]):\n", " print_colored(f\"\\nEmail {idx + 1} - Organizzazioni estratte:\", \"blue\")\n", " print({', '.join(orgs) if orgs else 'Nessuna'})\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Raggruppamento dei contenuti delle email

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Raggruppamento dei contenuti delle email NON SPAM in un DataFrame ordinato per frequenza\n\n           Organizzazione  Frequenza              Tipo         Fonti\n0                   enron        580  Da categorizzare  Da compilare\n1                      cc        387  Da categorizzare  Da compilare\n2                     hpl        273  Da categorizzare  Da compilare\n3                  sitara        249  Da categorizzare  Da compilare\n4                     ect        240  Da categorizzare  Da compilare\n...                   ...        ...               ...           ...\n3154  pioneer exploration          1  Da categorizzare  Da compilare\n3155               maytag          1  Da categorizzare  Da compilare\n3156                sears          1  Da categorizzare  Da compilare\n3157          sears . com          1  Da categorizzare  Da compilare\n3158                   ft          1  Da categorizzare  Da compilare\n\n[3159 rows x 4 columns]\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Raggruppamento delle organizzazioni in una lista unica\n", "flattened_organizations = [org for orgs in all_organizations for org in orgs]\n", "\n", "# Conto le occorrenze di ogni organizzazione\n", "org_counter = Counter(flattened_organizations)\n", "\n", "df_organizations = pd.DataFrame.from_dict(org_counter, orient='index', columns=['Frequenza']).reset_index()\n", "df_organizations.rename(columns={'index': 'Organizzazione'}, inplace=True)\n", "\n", "df_organizations['Tipo'] = 'Da categorizzare'\n", "df_organizations['Fonti'] = 'Da compilare'\n", "\n", "df_organizations = df_organizations.sort_values(by='Frequenza', ascending=False).reset_index(drop=True)\n", "\n", "print_colored(\"Raggruppamento dei contenuti delle email NON SPAM in un DataFrame ordinato per frequenza\\n\", \"blue\")\n", "print(df_organizations)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Raggruppamento dei contenuti delle email NON SPAM in un DataFrame ordinato per frequenza\n\n                          Organizzazione  Frequenza              Tipo  \\\n0                                  enron        580  Da categorizzare   \n1                                     cc        387  Da categorizzare   \n2                                    hpl        273  Da categorizzare   \n3                                 sitara        249  Da categorizzare   \n4                                    ect        240  Da categorizzare   \n5                                    ena        142  Da categorizzare   \n6                               eastrans         96  Da categorizzare   \n7                                     tu         94  Da categorizzare   \n8                                   hplc         91  Da categorizzare   \n9                        texas utilities         86  Da categorizzare   \n10                                 unify         86  Da categorizzare   \n11                           enron . com         80  Da categorizzare   \n12              enron north america corp         80  Da categorizzare   \n13                            tenaska iv         76  Da categorizzare   \n14                            enronxgate         63  Da categorizzare   \n15                              wellhead         48  Da categorizzare   \n16                                  duke         46  Da categorizzare   \n17                                 entex         44  Da categorizzare   \n18                              equistar         43  Da categorizzare   \n19                               tenaska         43  Da categorizzare   \n20                                   txu         41  Da categorizzare   \n21                                  deal         39  Da categorizzare   \n22                                  pops         39  Da categorizzare   \n23                                 tufco         36  Da categorizzare   \n24                               enserch         35  Da categorizzare   \n25                                  teco         35  Da categorizzare   \n26  enron capital & trade resources corp         34  Da categorizzare   \n27                                midcon         33  Da categorizzare   \n28                                   aep         32  Da categorizzare   \n29                                   gas         32  Da categorizzare   \n30                    north america corp         31  Da categorizzare   \n31                                   pec         31  Da categorizzare   \n32                   fuels cotton valley         30  Da categorizzare   \n33                                   ces         29  Da categorizzare   \n34                                pg & e         29  Da categorizzare   \n\n           Fonti  \n0   Da compilare  \n1   Da compilare  \n2   Da compilare  \n3   Da compilare  \n4   Da compilare  \n5   Da compilare  \n6   Da compilare  \n7   Da compilare  \n8   Da compilare  \n9   Da compilare  \n10  Da compilare  \n11  Da compilare  \n12  Da compilare  \n13  Da compilare  \n14  Da compilare  \n15  Da compilare  \n16  Da compilare  \n17  Da compilare  \n18  Da compilare  \n19  Da compilare  \n20  Da compilare  \n21  Da compilare  \n22  Da compilare  \n23  Da compilare  \n24  Da compilare  \n25  Da compilare  \n26  Da compilare  \n27  Da compilare  \n28  Da compilare  \n29  Da compilare  \n30  Da compilare  \n31  Da compilare  \n32  Da compilare  \n33  Da compilare  \n34  Da compilare  \n
\n
\n
\n
\n
" ] } } ], "source": [ "# Raggruppamento delle organizzazioni in una lista unica\n", "flattened_organizations = [org for orgs in all_organizations for org in orgs]\n", "\n", "org_counter = Counter(flattened_organizations)\n", "\n", "df_organizations = pd.DataFrame.from_dict(org_counter, orient='index', columns=['Frequenza']).reset_index()\n", "df_organizations.rename(columns={'index': 'Organizzazione'}, inplace=True)\n", "\n", "df_organizations['Tipo'] = 'Da categorizzare'\n", "df_organizations['Fonti'] = 'Da compilare'\n", "\n", "df_organizations = df_organizations.sort_values(by='Frequenza', ascending=False).reset_index(drop=True)\n", "\n", "# Configuro Pandas per non troncare le righe\n", "pd.set_option('display.max_rows', None)\n", "\n", "print_colored(\"Raggruppamento dei contenuti delle email NON SPAM in un DataFrame ordinato per frequenza\\n\", \"blue\")\n", "print(df_organizations.iloc[0:35])\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Scatter plot con riduzione della dimensionalità tramite t-SNE

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Le dimensioni delle bolle sono proporzionali alla frequenza delle organizzazioni

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Rappresentazione delle prime 35 Organizzazioni con 5 cluster in due dimensioni

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "from sklearn.feature_extraction.text import TfidfVectorizer\n", "\n", "# Dati iniziali\n", "orgs = df_organizations['Organizzazione'].values\n", "frequenze = df_organizations['Frequenza'].values\n", "\n", "# Converto i dati in rappresentazioni numeriche\n", "vectorizer = TfidfVectorizer()\n", "X = vectorizer.fit_transform(orgs)\n", "\n", "# Applico clustering \n", "n_clusters = 5 # Numero di cluster desiderati\n", "kmeans = KMeans(n_clusters=n_clusters, random_state=42)\n", "clusters = kmeans.fit_predict(X)\n", "\n", "# Riduzione dimensionale per la visualizzazione\n", "tsne = TSNE(n_components=2, random_state=42)\n", "X_embedded = tsne.fit_transform(X.toarray())\n", "\n", "df_organizations['Cluster'] = clusters\n", "\n", "cluster_names = []\n", "for cluster_id in range(n_clusters):\n", " cluster_orgs = df_organizations[df_organizations['Cluster'] == cluster_id].sort_values(by='Frequenza', ascending=False)['Organizzazione']\n", " top_keywords = \", \".join(cluster_orgs.head(35)) # Prime 35 organizzazioni come rappresentanti\n", " cluster_names.append(f\"Cluster {cluster_id}: {top_keywords}\")\n", "\n", "# Definizione dei colori delle bolle\n", "colors = plt.cm.get_cmap(\"tab10\", n_clusters).colors\n", "\n", "# Visualizzo i risultati con un grafico scatter\n", "plt.figure(figsize=(40, 40))\n", "scatter_points = [] # Per gestire i punti scatter\n", "scatter_labels = [] # Per gestire le etichette delle legende\n", "\n", "# Creo i grafici scatter\n", "for cluster_id in range(n_clusters):\n", " # Filtro i punti del cluster corrente\n", " cluster_mask = clusters == cluster_id\n", " cluster_points = X_embedded[cluster_mask]\n", " \n", " scatter = plt.scatter(\n", " cluster_points[:, 0], \n", " cluster_points[:, 1], \n", " s=frequenze[cluster_mask] * 70, # Dimensioni proporzionali alla frequenza\n", " c=[colors[cluster_id]], \n", " alpha=0.7, \n", " edgecolors='black'\n", " )\n", " scatter_points.append(scatter)\n", " scatter_labels.append(cluster_names[cluster_id])\n", "\n", "plt.title(\"\\nRappresentazione delle prime 35 Organizzazioni con 5 cluster\", fontsize=62)\n", "plt.xlabel(\"\\nt-SNE Dimension x\", fontsize=52)\n", "plt.ylabel(\"t-SNE Dimension y\", fontsize=52)\n", "plt.xticks(ha='right', fontsize=46, color='#b81414')\n", "plt.yticks(ha='right', fontsize=46, color='#b81414')\n", "plt.grid(True)\n", "\n", "plt.tight_layout()\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Rappresentazione delle prime 35 Organizzazioni con 20 cluster in due dimensioni

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "orgs = df_organizations['Organizzazione'].values\n", "frequenze = df_organizations['Frequenza'].values\n", "\n", "vectorizer = TfidfVectorizer()\n", "X = vectorizer.fit_transform(orgs)\n", "\n", "n_clusters = 20 \n", "kmeans = KMeans(n_clusters=n_clusters, random_state=42)\n", "clusters = kmeans.fit_predict(X)\n", "\n", "tsne = TSNE(n_components=2, random_state=42)\n", "X_embedded = tsne.fit_transform(X.toarray())\n", "\n", "df_organizations['Cluster'] = clusters\n", "\n", "cluster_names = []\n", "for cluster_id in range(n_clusters):\n", " cluster_orgs = df_organizations[df_organizations['Cluster'] == cluster_id].sort_values(by='Frequenza', ascending=False)['Organizzazione']\n", " top_keywords = \", \".join(cluster_orgs.head(35)) # Prime 35 organizzazioni come rappresentanti\n", " cluster_names.append(f\"Cluster {cluster_id}: {top_keywords}\")\n", "\n", "colors = plt.cm.get_cmap(\"tab10\", n_clusters).colors\n", "\n", "plt.figure(figsize=(40, 40))\n", "scatter_points = [] # Per gestire i punti scatter\n", "scatter_labels = [] # Per gestire le etichette delle legende\n", "\n", "for cluster_id in range(n_clusters):\n", " # Filtro i punti del cluster corrente\n", " cluster_mask = clusters == cluster_id\n", " cluster_points = X_embedded[cluster_mask]\n", " \n", " scatter = plt.scatter(\n", " cluster_points[:, 0], \n", " cluster_points[:, 1], \n", " s=frequenze[cluster_mask] * 70, \n", " c=[colors[cluster_id]], \n", " alpha=0.7, \n", " edgecolors='black'\n", " )\n", " scatter_points.append(scatter)\n", " scatter_labels.append(cluster_names[cluster_id])\n", "\n", "plt.title(\"\\nRappresentazione delle prime 35 Organizzazioni con 20 cluster\", fontsize=62)\n", "plt.xlabel(\"\\nt-SNE Dimension x\", fontsize=52)\n", "plt.ylabel(\"t-SNE Dimension y\", fontsize=52)\n", "plt.xticks(ha='right', fontsize=46, color='#b81414')\n", "plt.yticks(ha='right', fontsize=46, color='#b81414')\n", "plt.grid(True)\n", "\n", "plt.tight_layout()\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Rappresentando un numero maggiore di cluster in uno scatter plot con riduzione della dimensionalità tramite t-SNE, si ottiene una visione più dettagliata dell'eterogeneità dei dati.

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Rappresentazione delle prime 35 Organizzazioni con 20 cluster in 3 dimensioni

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "orgs = df_organizations['Organizzazione'].values\n", "frequenze = df_organizations['Frequenza'].values\n", "\n", "vectorizer = TfidfVectorizer()\n", "X = vectorizer.fit_transform(orgs)\n", "\n", "n_clusters = 20 \n", "kmeans = KMeans(n_clusters=n_clusters, random_state=42)\n", "clusters = kmeans.fit_predict(X)\n", "\n", "tsne = TSNE(n_components=3, random_state=42)\n", "X_embedded = tsne.fit_transform(X.toarray())\n", "\n", "df_organizations['Cluster'] = clusters\n", "\n", "cluster_names = []\n", "for cluster_id in range(n_clusters):\n", " cluster_orgs = df_organizations[df_organizations['Cluster'] == cluster_id].sort_values(by='Frequenza', ascending=False)['Organizzazione']\n", " top_keywords = \", \".join(cluster_orgs.head(35)) # Prime 35 organizzazioni come rappresentanti\n", " cluster_names.append(f\"Cluster {cluster_id}: {top_keywords}\")\n", "\n", "colors = plt.cm.get_cmap(\"tab10\", n_clusters).colors\n", "\n", "plt.figure(figsize=(40, 40))\n", "scatter_points = [] \n", "scatter_labels = [] \n", "\n", "for cluster_id in range(n_clusters):\n", "\n", " cluster_mask = clusters == cluster_id\n", " cluster_points = X_embedded[cluster_mask]\n", "\n", " scatter = plt.scatter(\n", " cluster_points[:, 0], \n", " cluster_points[:, 1], \n", " s=frequenze[cluster_mask] * 70, \n", " c=[colors[cluster_id]], \n", " alpha=0.7, \n", " edgecolors='black'\n", " )\n", " scatter_points.append(scatter)\n", " scatter_labels.append(cluster_names[cluster_id])\n", "\n", "plt.title(\"\\nRappresentazione delle prime 35 Organizzazioni con 20 cluster\", fontsize=62)\n", "plt.xlabel(\"\\nt-SNE Dimension x\", fontsize=52)\n", "plt.ylabel(\"t-SNE Dimension y\", fontsize=52)\n", "plt.xticks(ha='right', fontsize=46, color='#b81414')\n", "plt.yticks(ha='right', fontsize=46, color='#b81414')\n", "plt.grid(True)\n", "\n", "plt.tight_layout()\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Rappresentazione delle prime 35 Organizzazioni con 20 cluster in 3D

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "orgs = df_organizations['Organizzazione'].values\n", "frequenze = df_organizations['Frequenza'].values\n", "\n", "vectorizer = TfidfVectorizer()\n", "X = vectorizer.fit_transform(orgs)\n", "\n", "n_clusters = 20 \n", "kmeans = KMeans(n_clusters=n_clusters, random_state=42)\n", "clusters = kmeans.fit_predict(X)\n", "\n", "tsne = TSNE(n_components=3, random_state=42)\n", "X_embedded = tsne.fit_transform(X.toarray())\n", "\n", "df_organizations['Cluster'] = clusters\n", "\n", "cluster_names = []\n", "for cluster_id in range(n_clusters):\n", " cluster_orgs = df_organizations[df_organizations['Cluster'] == cluster_id].sort_values(by='Frequenza', ascending=False)['Organizzazione']\n", " top_keywords = \", \".join(cluster_orgs.head(35)) # Prime 35 organizzazioni come rappresentanti\n", " cluster_names.append(f\"Cluster {cluster_id}: {top_keywords}\")\n", "\n", "colors = plt.cm.get_cmap(\"tab10\", n_clusters).colors\n", "\n", "fig = plt.figure(figsize=(40, 40))\n", "ax = fig.add_subplot(111, projection='3d') \n", "\n", "scatter_points = [] \n", "scatter_labels = [] \n", "\n", "# scatter 3D\n", "for cluster_id in range(n_clusters):\n", " \n", " cluster_mask = clusters == cluster_id\n", " cluster_points = X_embedded[cluster_mask]\n", " \n", " scatter = ax.scatter(\n", " cluster_points[:, 0], \n", " cluster_points[:, 1], \n", " cluster_points[:, 2], \n", " s=frequenze[cluster_mask] * 70, \n", " c=[colors[cluster_id]], \n", " alpha=0.7, \n", " edgecolors='black'\n", " )\n", " scatter_points.append(scatter)\n", " scatter_labels.append(cluster_names[cluster_id])\n", "\n", "ax.set_title(\"\\nRappresentazione delle prime 35 Organizzazioni con 20 cluster\", fontsize=62)\n", "ax.set_xlabel(\"t-SNE Dimension x\", fontsize=52, labelpad=40)\n", "ax.set_ylabel(\"t-SNE Dimension y\", fontsize=52, labelpad=45)\n", "ax.set_zlabel(\"t-SNE Dimension z\", fontsize=52, labelpad=40)\n", "ax.view_init(elev=30, azim=45) \n", "\n", "ax.tick_params(axis='x', labelsize=46, colors='#b81414')\n", "ax.tick_params(axis='y', labelsize=46, colors='#b81414')\n", "ax.tick_params(axis='z', labelsize=46, colors='#b81414')\n", "plt.grid(True)\n", "\n", "plt.tight_layout()\n", "\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Grafico 3D interattivo

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
" ] } } ], "source": [ "import plotly.express as px\n", "\n", "orgs = df_organizations['Organizzazione'].values\n", "frequenze = df_organizations['Frequenza'].values\n", "\n", "vectorizer = TfidfVectorizer()\n", "X = vectorizer.fit_transform(orgs)\n", "\n", "n_clusters = 20 \n", "kmeans = KMeans(n_clusters=n_clusters, random_state=42)\n", "clusters = kmeans.fit_predict(X)\n", "\n", "tsne = TSNE(n_components=3, random_state=42, perplexity=10)\n", "X_embedded = tsne.fit_transform(X.toarray())\n", "\n", "df_organizations['Cluster'] = clusters\n", "\n", "cluster_names = []\n", "for cluster_id in range(n_clusters):\n", " cluster_orgs = df_organizations[df_organizations['Cluster'] == cluster_id].sort_values(by='Frequenza', ascending=False)['Organizzazione']\n", " top_keywords = \", \".join(cluster_orgs.head(35)) \n", " cluster_names.append(f\"Cluster {cluster_id}: {top_keywords}\")\n", "\n", "colors = px.colors.qualitative.Set2 \n", "\n", "df_vis = pd.DataFrame(X_embedded, columns=['Dim1', 'Dim2', 'Dim3'])\n", "df_vis['Organizzazione'] = orgs\n", "df_vis['Frequenza'] = frequenze\n", "df_vis['Cluster'] = clusters.astype(str) # Converto i cluster in stringa per la legenda\n", "\n", "# Normalizzo le dimensioni per evitare valori troppo piccoli o nulli\n", "df_vis['Frequenza'] = df_vis['Frequenza'].fillna(1) # Rimpiazza eventuali valori NaN con 1\n", "df_vis['Size'] = df_vis['Frequenza'] * 20000000000000 # Scala della dimensione delle bolle\n", "\n", "\"\"\"\n", "Purtroppo le bolle sono sempre piccolissime\n", "\"\"\"\n", "\n", "# Grafico scatter 3D con Plotly\n", "fig = px.scatter_3d(\n", " df_vis, \n", " x='Dim1', \n", " y='Dim2', \n", " z='Dim3', \n", " color='Cluster', # Cluster come colori\n", " size='Size', # Dimensioni proporzionali alla frequenza\n", " hover_name='Organizzazione', \n", " title=\"Rappresentazione delle prime 35 Organizzazioni con 20 cluster (3D)\",\n", " labels={'Dim1': 't-SNE Dimension x', 'Dim2': 't-SNE Dimension y', 'Dim3': 't-SNE Dimension z'},\n", " color_discrete_sequence=colors \n", ")\n", "\n", "# Rimuovo il contorno nero e imposto colori pieni\n", "fig.update_traces(marker=dict(line=dict(width=0))) \n", "\n", "fig.update_layout(\n", " width=1600, \n", " height=1000,\n", " legend_title=\"Cluster\",\n", " scene=dict(\n", " xaxis_title=\"t-SNE Dimension x\",\n", " yaxis_title=\"t-SNE Dimension y\",\n", " zaxis_title=\"t-SNE Dimension z\"\n", " )\n", ")\n", "\n", "fig.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Grafico a bolle

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Le dimensioni delle bolle rappresentano la frequenza delle organizzazioni.

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "import itertools\n", "\n", "top_organizations = df_organizations.nlargest(35, 'Frequenza')\n", "\n", "x = top_organizations['Organizzazione']\n", "y = top_organizations['Frequenza']\n", "sizes = top_organizations['Frequenza'] * 10 # Scala delle dimensioni delle bolle\n", "\n", "colors = ['#0897B4', '#0B2B40', '#FF5F5D'] \n", "color_cycle = list(itertools.islice(itertools.cycle(colors), len(x)))\n", "\n", "plt.figure(figsize=(12, 8))\n", "plt.scatter(x, y, s=sizes, c=color_cycle, edgecolors='black', linewidth=0.3)\n", "\n", "plt.xticks(\n", " ticks=range(len(x)), # Indici delle etichette\n", " labels=[f\"{label}\" for label in x], # Etichette\n", " rotation=45, \n", " ha='right', # Allineamento\n", " fontsize=10, \n", " color='black' \n", ")\n", "\n", "# Coloro individualmente le etichette\n", "for i, tick_label in enumerate(plt.gca().get_xticklabels()):\n", " tick_label.set_color(color_cycle[i]) # Colore corrispondente alla bolla\n", "\n", "plt.title('\\nRappresentazione delle prime 35 Organizzazioni', fontsize=18)\n", "plt.xlabel('Organizzazioni', fontsize=16)\n", "plt.ylabel('Frequenza', fontsize=16)\n", "plt.yticks(fontsize=14, color='#b81414')\n", "\n", "plt.tight_layout()\n", "\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Costruzione e addestramento di un classificatore per la rilevazione di email SPAM

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Analisi del dataset

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Breve recap sul dataset raw iniziale

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
Out[34]:
\n
\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Unnamed: 0labeltextlabel_num
0605hamSubject: enron methanol ; meter # : 988291\\nth...0
12349hamSubject: hpl nom for january 9 , 2001\\n( see a...0
23624hamSubject: neon retreat\\nho ho ho , we ' re arou...0
34685spamSubject: photoshop , windows , office . cheap ...1
42030hamSubject: re : indian springs\\nthis deal is to ...0
\n
\n
\n
\n
\n
" ] } } ], "source": [ "df_email.head()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Il peso del dataset raw in memoria è di:\n5.66 MB\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Calcolo del peso totale del dataset in memoria\n", "df_email_size = df_email.memory_usage(deep=True).sum()\n", "\n", "# Conversione in megabyte (MB)\n", "df_email_size_mb = df_email_size / (1024 ** 2)\n", "\n", "print_colored(f\"Il peso del dataset raw in memoria è di:\", \"blue\") \n", "print(f\"{df_email_size_mb:.2f} MB\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Dataset df_email dropped\n\n                                                text  label_num\n0  Subject: enron methanol ; meter # : 988291\\nth...          0\n1  Subject: hpl nom for january 9 , 2001\\n( see a...          0\n2  Subject: neon retreat\\nho ho ho , we ' re arou...          0\n3  Subject: photoshop , windows , office . cheap ...          1\n4  Subject: re : indian springs\\nthis deal is to ...          0\n
\n
\n
\n
\n
" ] } } ], "source": [ "df_email_dropped = df_email.drop(columns=['Unnamed: 0', 'label'])\n", "\n", "print_colored(\"Dataset df_email dropped\\n\", \"blue\")\n", "print(df_email_dropped.head())\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "label_count.plot(kind='bar', figsize=(10, 6))\n", "plt.title(\"\\nDistribuzione di ham e spam\", fontsize=18)\n", "plt.xlabel(\"Label (0 = ham, 1 = spam)\", fontsize=14)\n", "plt.ylabel(\"Conteggio\", fontsize=14)\n", "plt.xticks(fontsize=12, color='#b81414', rotation=0)\n", "plt.yticks(fontsize=12, color='#b81414')\n", "plt.grid(axis='y', linestyle='--', alpha=0.5, color=\"#1f77b4\")\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Preprocessing del testo delle email HAM e SPAM

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Faccio il preprocessing prima del bilanciamento del dataset per evitare di duplicare dati sporchi, per avere una migliore qualità dei dati e per ridurre i calcoli computazionali

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Colonna 'text' originale\n                                                text\n0  Subject: enron methanol ; meter # : 988291\\nth...\n1  Subject: hpl nom for january 9 , 2001\\n( see a...\n2  Subject: neon retreat\\nho ho ho , we ' re arou...\n3  Subject: photoshop , windows , office . cheap ...\n4  Subject: re : indian springs\\nthis deal is to ...\n\nColonna 'clean_text' con l'applicazione della funzione 'clean_text' alla colonna 'text'\n                                        cleaned_text\n0  enron methanol meter follow note give monday p...\n1  hpl nom january see attach file hplnol xls hpl...\n2  neon retreat ho ho ho around wonderful time ye...\n3  photoshop windows office cheap main trend abas...\n4  indian spring deal book teco pvr revenue under...\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Inizializzo stopwords e lemmatizer\n", "stop_words = set(stopwords.words('english')) # Stopword per la lingua inglese\n", " # Rimuove parole comuni inglesi per ridurre il rumore nei dati\n", "lemmatizer = WordNetLemmatizer() # Riduce le parole alla loro forma base per \n", " # diminuire la dimensionalità del vocabolario\n", "\n", "def clean_text(text):\n", " # Rimuovo caratteri speciali e punteggiatura\n", " text = re.sub(r'[^a-zA-Z\\s]', '', text) \n", " \n", " # Converto tutte le lettere in minuscolo\n", " text = text.lower()\n", " \n", " # Rimuovo la parola 'subject' se presente perché si riferisce alla struttura standard di un'email\n", " # non aggiunge valore informativo per l'analisi, anzi, può influenzarla negativamente\n", " text = text.replace('subject', '') \n", " \n", " tokens = [lemmatizer.lemmatize(word, pos='v') for word in text.split() if word not in stop_words]\n", " return ' '.join(tokens)\n", "\n", "# Creo la colonna cleaned_text e applico la pulizia del testo\n", "df_email_dropped['cleaned_text'] = df_email_dropped['text'].apply(clean_text)\n", "\n", "print_colored(\"Colonna 'text' originale\", \"blue\")\n", "print(df_email_dropped[['text']].head())\n", "\n", "print()\n", "\n", "print_colored(\"Colonna 'clean_text' con l'applicazione della funzione 'clean_text' alla colonna 'text'\", \"blue\")\n", "print(df_email_dropped[['cleaned_text']].head())\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
Out[39]:
\n
\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
textlabel_numcleaned_text
0Subject: enron methanol ; meter # : 988291\\nth...0enron methanol meter follow note give monday p...
1Subject: hpl nom for january 9 , 2001\\n( see a...0hpl nom january see attach file hplnol xls hpl...
2Subject: neon retreat\\nho ho ho , we ' re arou...0neon retreat ho ho ho around wonderful time ye...
3Subject: photoshop , windows , office . cheap ...1photoshop windows office cheap main trend abas...
4Subject: re : indian springs\\nthis deal is to ...0indian spring deal book teco pvr revenue under...
\n
\n
\n
\n
\n
" ] } } ], "source": [ "df_email_dropped.head()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Il peso del dataset 'df_email_dropped' in memoria è di:\n8.68 MB\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Calcolo del peso totale del dataset in memoria\n", "df_email_dropped_size = df_email_dropped.memory_usage(deep=True).sum()\n", "\n", "# Conversione in megabyte (MB)\n", "df_email_dropped_mb = df_email_dropped_size / (1024 ** 2)\n", "\n", "print_colored(f\"Il peso del dataset 'df_email_dropped' in memoria è di:\", \"blue\") \n", "print(f\"{df_email_dropped_mb:.2f} MB\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Dataset 'df_email_dropped' cleaned\n\n   label_num                                       cleaned_text\n0          0  enron methanol meter follow note give monday p...\n1          0  hpl nom january see attach file hplnol xls hpl...\n2          0  neon retreat ho ho ho around wonderful time ye...\n3          1  photoshop windows office cheap main trend abas...\n4          0  indian spring deal book teco pvr revenue under...\n
\n
\n
\n
\n
" ] } } ], "source": [ "df_email_dropped_cleaned = df_email_dropped.drop(columns=['text'])\n", "\n", "print_colored(\"Dataset 'df_email_dropped' cleaned\\n\", \"blue\")\n", "print(df_email_dropped_cleaned.head())\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Il peso del dataset raw in memoria è di:\n3.36 MB\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Calcolo del peso totale del dataset in memoria\n", "df_email_dropped_cleaned_size = df_email_dropped_cleaned.memory_usage(deep=True).sum()\n", "\n", "# Conversione in megabyte (MB)\n", "df_email_dropped_cleaned_mb = df_email_dropped_cleaned_size / (1024 ** 2)\n", "\n", "print_colored(f\"Il peso del dataset raw in memoria è di:\", \"blue\") \n", "print(f\"{df_email_dropped_cleaned_mb:.2f} MB\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
La memoria liberata grazie al dropout e alla pulizia del testo rispetto al dataset di partenza è di:\n2.30 MB\nPercentuale di memoria liberata:\n25.54 %\n
\n
\n
\n
\n
" ] } } ], "source": [ "dataset_size_difference = df_email_size_mb - df_email_dropped_cleaned_mb\n", "\n", "print_colored(f\"La memoria liberata grazie al dropout e alla pulizia del testo rispetto al dataset di partenza è di:\", \"blue\") \n", "print(f\"{dataset_size_difference:.2f} MB\")\n", "\n", "total_size = df_email_size_mb + df_email_dropped_cleaned_mb\n", "difference_percentage_size = (dataset_size_difference/total_size)*100\n", "\n", "print_colored(f\"Percentuale di memoria liberata:\", \"blue\") \n", "print(f\"{difference_percentage_size:.2f} %\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Implemento il bilanciamento delle classi con SMOTE

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Distribuzione delle classi dopo SMOTE:\nCounter({0: 3672, 1: 3672})\n
\n
\n
\n
\n
" ] } } ], "source": [ "from imblearn.over_sampling import SMOTE\n", "\n", "# Estraggo feature e label\n", "X = df_email_dropped_cleaned['cleaned_text']\n", "y = df_email_dropped_cleaned['label_num']\n", "\n", "# Conversione del testo in rappresentazione numerica TF-IDF\n", "vectorizer = TfidfVectorizer(max_features=5000) # Limito il numero di feature per evitare sovraccarico di memoria\n", "X_tfidf = vectorizer.fit_transform(X)\n", "\n", "# Converto in DataFrame se necessario\n", "if not isinstance(X_tfidf, pd.DataFrame):\n", " X_tfidf = pd.DataFrame(X_tfidf.toarray(), columns=[f'feature_{i}' for i in range(X_tfidf.shape[1])])\n", "if not isinstance(y, pd.DataFrame):\n", " y = pd.DataFrame(y, columns=['label_num'])\n", "\n", "# Applico SMOTE per bilanciare le classi\n", "smote = SMOTE(random_state=42)\n", "X_resampled, y_resampled = smote.fit_resample(X_tfidf, y)\n", "\n", "# Converto i dati bilanciati in DataFrame per facilitarne l'uso successivo\n", "df_balanced = pd.concat([pd.DataFrame(X_resampled, columns=X_tfidf.columns), pd.DataFrame(y_resampled, columns=['label_num'])], axis=1)\n", "\n", "print_colored(\"Distribuzione delle classi dopo SMOTE:\", \"blue\")\n", "print(Counter(y_resampled['label_num']))\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "label_counts = df_balanced['label_num'].value_counts()\n", "\n", "df_plot = pd.DataFrame({\n", " 'Categoria': ['Ham', 'Spam'],\n", " 'Conteggio': [label_counts.get(0, 0), label_counts.get(1, 0)]\n", "})\n", "\n", "plt.figure(figsize=(10, 6))\n", "plt.bar(df_plot['Categoria'], df_plot['Conteggio'], color=['#1f77b4', '#ff7f0e'])\n", "\n", "plt.title(\"\\nDistribuzione di ham e spam\", fontsize=18)\n", "plt.xlabel(\"Categoria\", fontsize=14)\n", "plt.ylabel(\"Conteggio\", fontsize=14)\n", "plt.xticks(fontsize=12, color='#b81414', rotation=0)\n", "plt.yticks(fontsize=12, color='#b81414')\n", "plt.grid(axis='y', linestyle='--', alpha=0.5, color=\"#1f77b4\")\n", "\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Creazione dei classificatori

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Training set:\n3619\n\nVal set:\n776\n\nTest set:\n776\n
\n
\n
\n
\n
" ] } } ], "source": [ "from tensorflow.keras.preprocessing.sequence import pad_sequences\n", "from tensorflow.keras.models import Sequential\n", "from tensorflow.keras.layers import Embedding, Conv1D, MaxPooling1D, LSTM, GRU, Dense, Dropout, Bidirectional, Flatten\n", "from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau\n", "from sklearn.model_selection import train_test_split\n", "from sklearn.feature_extraction.text import TfidfVectorizer\n", "from tensorflow.keras.preprocessing.text import Tokenizer\n", "\n", "# Suddivisione del dataset in train, validation e test (70-15-15) con stratificazione\n", "X = df_email_dropped_cleaned['cleaned_text'].values # Testi puliti\n", "y = df_email_dropped_cleaned['label_num'].values # Etichette spam/ham\n", "\n", "X_train, X_temp, y_train, y_temp = train_test_split(X, y, test_size=0.30, stratify=y, random_state=42)\n", "X_val, X_test, y_val, y_test = train_test_split(X_temp, y_temp, test_size=0.50, stratify=y_temp, random_state=42)\n", "\n", "print_colored(f\"Training set:\", \"blue\") \n", "print(len(X_train))\n", "\n", "print_colored(f\"\\nVal set:\", \"blue\") \n", "print(len(X_val))\n", "\n", "print_colored(f\"\\nTest set:\", \"blue\") \n", "print(len(X_test))\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Conversione dei dati per l'uso con reti neurali (TensorFlow)\n", "X_train_tfidf = tf.convert_to_tensor(X_train_tfidf, dtype=tf.float32)\n", "X_val_tfidf = tf.convert_to_tensor(X_val_tfidf, dtype=tf.float32)\n", "X_test_tfidf = tf.convert_to_tensor(X_test_tfidf, dtype=tf.float32)\n", "\n", "y_train = tf.convert_to_tensor(y_train, dtype=tf.float32)\n", "y_val = tf.convert_to_tensor(y_val, dtype=tf.float32)\n", "y_test = tf.convert_to_tensor(y_test, dtype=tf.float32)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Shape X_train: (3619, 3500) Y_train: (3619,)\nShape X_val: (776, 3500) Y_val: (776,)\nShape X_test: (776, 3500) Y_test: (776,)\n
\n
\n
\n
\n
Out[50]:
\n
\n
['classificatori/MLP/tfidf_vectorizer.pkl']
\n
\n
\n
\n
" ] } } ], "source": [ "# Verifico la dimensione dei dati\n", "print(\"Shape X_train:\", X_train_tfidf.shape, \"Y_train:\", y_train.shape)\n", "print(\"Shape X_val:\", X_val_tfidf.shape, \"Y_val:\", y_val.shape)\n", "print(\"Shape X_test:\", X_test_tfidf.shape, \"Y_test:\", y_test.shape)\n", "\n", "import joblib\n", "joblib.dump(vectorizer, 'classificatori/MLP/tfidf_vectorizer.pkl')\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Modello MLP

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
MLP\n
\n
\n
\n
\n
" ] } } ], "source": [ "model_name = \"MLP\"\n", "print(model_name)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Callbacks per il monitoraggio\n", "callbacks = [\n", " EarlyStopping(\n", " monitor='val_loss', \n", " patience=10, \n", " restore_best_weights=True, \n", " mode='min',\n", " min_delta=0.001\n", " ), \n", " \n", " ReduceLROnPlateau(\n", " monitor='val_loss', \n", " factor=0.5, \n", " patience=5, \n", " min_lr=1e-6,\n", " verbose=1,\n", " ),\n", " \n", " ModelCheckpoint(\n", " f\"classificatori/MLP/{model_name}_best_model.keras\",\n", " monitor='val_loss', \n", " save_best_only=True, \n", " verbose=1\n", " )\n", "]\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from tensorflow.keras.regularizers import l1, l2\n", "\n", "mlp_model = Sequential([\n", " Dense(256, activation='relu', input_shape=(X_train_tfidf.shape[1],),\n", " kernel_regularizer=l2(0.001)), \n", " Dropout(0.3), \n", " Dense(128, activation='relu', \n", " kernel_regularizer=l2(0.001)),\n", " Dropout(0.3),\n", " Dense(64, activation='relu',\n", " kernel_regularizer=l2(0.001)),\n", " Dense(1, activation='sigmoid')\n", "])\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Model: \"sequential\"\n
\n
\n
\n
\n
\n
\n
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃\n┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n│ dense (Dense)                   │ (None, 256)            │       896,256 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ dropout (Dropout)               │ (None, 256)            │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ dense_1 (Dense)                 │ (None, 128)            │        32,896 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ dropout_1 (Dropout)             │ (None, 128)            │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ dense_2 (Dense)                 │ (None, 64)             │         8,256 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ dense_3 (Dense)                 │ (None, 1)              │            65 │\n└─────────────────────────────────┴────────────────────────┴───────────────┘\n
\n
\n
\n
\n
\n
\n
 Total params: 937,473 (3.58 MB)\n
\n
\n
\n
\n
\n
\n
 Trainable params: 937,473 (3.58 MB)\n
\n
\n
\n
\n
\n
\n
 Non-trainable params: 0 (0.00 B)\n
\n
\n
\n
\n
" ] } } ], "source": [ "mlp_model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0005), \n", " loss='binary_crossentropy', \n", " metrics=['accuracy'])\n", "\n", "mlp_model.summary()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Epoch 1/50\n
\n
\n
\n
\n
\n
\n
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR\nI0000 00:00:1737558883.542210   19651 service.cc:148] XLA service 0x7bce2c01d180 initialized for platform CUDA (this does not guarantee that XLA will be used). Devices:\nI0000 00:00:1737558883.542229   19651 service.cc:156]   StreamExecutor device (0): NVIDIA GeForce RTX 4070 Laptop GPU, Compute Capability 8.9\nI0000 00:00:1737558883.650929   19651 cuda_dnn.cc:529] Loaded cuDNN version 90600\n
\n
\n
\n
\n
\n
\n
 79/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.6899 - loss: 1.1224
\n
\n
\n
\n
\n
\n
I0000 00:00:1737558884.286870   19651 device_compiler.h:188] Compiled cluster using XLA!  This line is logged at most once for the lifetime of the process.\n
\n
\n
\n
\n
\n
\n
114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 10ms/step - accuracy: 0.7182 - loss: 1.0206\nEpoch 1: val_loss improved from inf to 0.31479, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 4s 25ms/step - accuracy: 0.7190 - loss: 1.0180 - val_accuracy: 0.9704 - val_loss: 0.3148 - learning_rate: 5.0000e-04\nEpoch 2/50\n103/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9890 - loss: 0.2398 \nEpoch 2: val_loss improved from 0.31479 to 0.18427, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9888 - loss: 0.2373 - val_accuracy: 0.9832 - val_loss: 0.1843 - learning_rate: 5.0000e-04\nEpoch 3/50\n 81/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9984 - loss: 0.1453 \nEpoch 3: val_loss improved from 0.18427 to 0.15297, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9982 - loss: 0.1429 - val_accuracy: 0.9820 - val_loss: 0.1530 - learning_rate: 5.0000e-04\nEpoch 4/50\n111/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9984 - loss: 0.1050 \nEpoch 4: val_loss improved from 0.15297 to 0.12904, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9984 - loss: 0.1049 - val_accuracy: 0.9768 - val_loss: 0.1290 - learning_rate: 5.0000e-04\nEpoch 5/50\n 79/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9995 - loss: 0.0915 \nEpoch 5: val_loss improved from 0.12904 to 0.11600, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9993 - loss: 0.0904 - val_accuracy: 0.9807 - val_loss: 0.1160 - learning_rate: 5.0000e-04\nEpoch 6/50\n 82/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9995 - loss: 0.0763 \nEpoch 6: val_loss improved from 0.11600 to 0.10993, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9994 - loss: 0.0759 - val_accuracy: 0.9781 - val_loss: 0.1099 - learning_rate: 5.0000e-04\nEpoch 7/50\n100/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9982 - loss: 0.0691 \nEpoch 7: val_loss improved from 0.10993 to 0.10971, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9983 - loss: 0.0692 - val_accuracy: 0.9794 - val_loss: 0.1097 - learning_rate: 5.0000e-04\nEpoch 8/50\n 81/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9985 - loss: 0.0677 \nEpoch 8: val_loss improved from 0.10971 to 0.10902, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9985 - loss: 0.0676 - val_accuracy: 0.9781 - val_loss: 0.1090 - learning_rate: 5.0000e-04\nEpoch 9/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9999 - loss: 0.0576 \nEpoch 9: val_loss improved from 0.10902 to 0.09849, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9999 - loss: 0.0576 - val_accuracy: 0.9794 - val_loss: 0.0985 - learning_rate: 5.0000e-04\nEpoch 10/50\n 83/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 1.0000 - loss: 0.0541 \nEpoch 10: val_loss did not improve from 0.09849\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 1.0000 - loss: 0.0541 - val_accuracy: 0.9742 - val_loss: 0.1081 - learning_rate: 5.0000e-04\nEpoch 11/50\n107/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9990 - loss: 0.0551 \nEpoch 11: val_loss did not improve from 0.09849\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - accuracy: 0.9990 - loss: 0.0552 - val_accuracy: 0.9742 - val_loss: 0.1010 - learning_rate: 5.0000e-04\nEpoch 12/50\n 82/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9991 - loss: 0.0517 \nEpoch 12: val_loss did not improve from 0.09849\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9989 - loss: 0.0527 - val_accuracy: 0.9807 - val_loss: 0.1046 - learning_rate: 5.0000e-04\nEpoch 13/50\n102/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9999 - loss: 0.0559 \nEpoch 13: val_loss improved from 0.09849 to 0.09842, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9998 - loss: 0.0560 - val_accuracy: 0.9794 - val_loss: 0.0984 - learning_rate: 5.0000e-04\nEpoch 14/50\n102/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9993 - loss: 0.0530 \nEpoch 14: val_loss improved from 0.09842 to 0.08891, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9994 - loss: 0.0528 - val_accuracy: 0.9820 - val_loss: 0.0889 - learning_rate: 5.0000e-04\nEpoch 15/50\n 78/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 1.0000 - loss: 0.0523 \nEpoch 15: val_loss did not improve from 0.08891\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9995 - loss: 0.0535 - val_accuracy: 0.9768 - val_loss: 0.1134 - learning_rate: 5.0000e-04\nEpoch 16/50\n 80/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9984 - loss: 0.0552 \nEpoch 16: val_loss did not improve from 0.08891\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9986 - loss: 0.0550 - val_accuracy: 0.9755 - val_loss: 0.1108 - learning_rate: 5.0000e-04\nEpoch 17/50\n 78/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9985 - loss: 0.0515 \nEpoch 17: val_loss did not improve from 0.08891\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9984 - loss: 0.0519 - val_accuracy: 0.9781 - val_loss: 0.1012 - learning_rate: 5.0000e-04\nEpoch 18/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 1.0000 - loss: 0.0485 \nEpoch 18: val_loss did not improve from 0.08891\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 1.0000 - loss: 0.0485 - val_accuracy: 0.9755 - val_loss: 0.1028 - learning_rate: 5.0000e-04\nEpoch 19/50\n110/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 1.0000 - loss: 0.0445 \nEpoch 19: val_loss improved from 0.08891 to 0.08301, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9999 - loss: 0.0445 - val_accuracy: 0.9845 - val_loss: 0.0830 - learning_rate: 5.0000e-04\nEpoch 20/50\n105/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9990 - loss: 0.0439 \nEpoch 20: val_loss did not improve from 0.08301\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9990 - loss: 0.0439 - val_accuracy: 0.9820 - val_loss: 0.0895 - learning_rate: 5.0000e-04\nEpoch 21/50\n 78/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9993 - loss: 0.0435 \nEpoch 21: val_loss did not improve from 0.08301\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9992 - loss: 0.0439 - val_accuracy: 0.9768 - val_loss: 0.0895 - learning_rate: 5.0000e-04\nEpoch 22/50\n100/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9997 - loss: 0.0425 \nEpoch 22: val_loss did not improve from 0.08301\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9997 - loss: 0.0426 - val_accuracy: 0.9820 - val_loss: 0.0842 - learning_rate: 5.0000e-04\nEpoch 23/50\n107/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9999 - loss: 0.0400 \nEpoch 23: val_loss did not improve from 0.08301\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9998 - loss: 0.0401 - val_accuracy: 0.9794 - val_loss: 0.0877 - learning_rate: 5.0000e-04\nEpoch 24/50\n 80/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9997 - loss: 0.0407 \nEpoch 24: ReduceLROnPlateau reducing learning rate to 0.0002500000118743628.\n\nEpoch 24: val_loss did not improve from 0.08301\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9994 - loss: 0.0415 - val_accuracy: 0.9832 - val_loss: 0.0905 - learning_rate: 5.0000e-04\nEpoch 25/50\n110/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9997 - loss: 0.0424 \nEpoch 25: val_loss improved from 0.08301 to 0.08101, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9997 - loss: 0.0424 - val_accuracy: 0.9820 - val_loss: 0.0810 - learning_rate: 2.5000e-04\nEpoch 26/50\n 79/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9984 - loss: 0.0398 \nEpoch 26: val_loss did not improve from 0.08101\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9987 - loss: 0.0391 - val_accuracy: 0.9794 - val_loss: 0.0812 - learning_rate: 2.5000e-04\nEpoch 27/50\n 82/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9998 - loss: 0.0346 \nEpoch 27: val_loss improved from 0.08101 to 0.07993, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9998 - loss: 0.0349 - val_accuracy: 0.9794 - val_loss: 0.0799 - learning_rate: 2.5000e-04\nEpoch 28/50\n 78/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9995 - loss: 0.0348 \nEpoch 28: val_loss improved from 0.07993 to 0.07549, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9995 - loss: 0.0347 - val_accuracy: 0.9832 - val_loss: 0.0755 - learning_rate: 2.5000e-04\nEpoch 29/50\n110/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9996 - loss: 0.0341 \nEpoch 29: val_loss did not improve from 0.07549\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9996 - loss: 0.0341 - val_accuracy: 0.9781 - val_loss: 0.0863 - learning_rate: 2.5000e-04\nEpoch 30/50\n110/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9996 - loss: 0.0342 \nEpoch 30: val_loss did not improve from 0.07549\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9996 - loss: 0.0342 - val_accuracy: 0.9820 - val_loss: 0.0759 - learning_rate: 2.5000e-04\nEpoch 31/50\n110/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 1.0000 - loss: 0.0315 \nEpoch 31: val_loss did not improve from 0.07549\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9999 - loss: 0.0316 - val_accuracy: 0.9832 - val_loss: 0.0768 - learning_rate: 2.5000e-04\nEpoch 32/50\n 80/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9991 - loss: 0.0345 \nEpoch 32: val_loss did not improve from 0.07549\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9993 - loss: 0.0341 - val_accuracy: 0.9820 - val_loss: 0.0760 - learning_rate: 2.5000e-04\nEpoch 33/50\n 81/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9997 - loss: 0.0325 \nEpoch 33: ReduceLROnPlateau reducing learning rate to 0.0001250000059371814.\n\nEpoch 33: val_loss did not improve from 0.07549\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9997 - loss: 0.0324 - val_accuracy: 0.9781 - val_loss: 0.0804 - learning_rate: 2.5000e-04\nEpoch 34/50\n 79/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 1.0000 - loss: 0.0298 \nEpoch 34: val_loss did not improve from 0.07549\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 1.0000 - loss: 0.0301 - val_accuracy: 0.9807 - val_loss: 0.0792 - learning_rate: 1.2500e-04\nEpoch 35/50\n 79/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9996 - loss: 0.0318 \nEpoch 35: val_loss did not improve from 0.07549\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9996 - loss: 0.0317 - val_accuracy: 0.9807 - val_loss: 0.0784 - learning_rate: 1.2500e-04\nEpoch 36/50\n 98/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 1.0000 - loss: 0.0295 \nEpoch 36: val_loss improved from 0.07549 to 0.07423, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 1.0000 - loss: 0.0296 - val_accuracy: 0.9820 - val_loss: 0.0742 - learning_rate: 1.2500e-04\nEpoch 37/50\n 80/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 1.0000 - loss: 0.0307 \nEpoch 37: val_loss did not improve from 0.07423\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 1.0000 - loss: 0.0306 - val_accuracy: 0.9768 - val_loss: 0.0755 - learning_rate: 1.2500e-04\nEpoch 38/50\n 80/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9989 - loss: 0.0312 \nEpoch 38: val_loss did not improve from 0.07423\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9991 - loss: 0.0308 - val_accuracy: 0.9807 - val_loss: 0.0749 - learning_rate: 1.2500e-04\nEpoch 39/50\n111/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9995 - loss: 0.0295 \nEpoch 39: val_loss did not improve from 0.07423\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9995 - loss: 0.0295 - val_accuracy: 0.9781 - val_loss: 0.0776 - learning_rate: 1.2500e-04\nEpoch 40/50\n 77/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9999 - loss: 0.0287 \nEpoch 40: val_loss did not improve from 0.07423\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9998 - loss: 0.0289 - val_accuracy: 0.9755 - val_loss: 0.0815 - learning_rate: 1.2500e-04\nEpoch 41/50\n 82/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9990 - loss: 0.0327 \nEpoch 41: ReduceLROnPlateau reducing learning rate to 6.25000029685907e-05.\n\nEpoch 41: val_loss did not improve from 0.07423\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9992 - loss: 0.0320 - val_accuracy: 0.9781 - val_loss: 0.0767 - learning_rate: 1.2500e-04\nEpoch 42/50\n 79/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 1.0000 - loss: 0.0282 \nEpoch 42: val_loss did not improve from 0.07423\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9999 - loss: 0.0286 - val_accuracy: 0.9807 - val_loss: 0.0749 - learning_rate: 6.2500e-05\nEpoch 43/50\n105/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9999 - loss: 0.0283 \nEpoch 43: val_loss improved from 0.07423 to 0.07366, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9999 - loss: 0.0284 - val_accuracy: 0.9807 - val_loss: 0.0737 - learning_rate: 6.2500e-05\nEpoch 44/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 1.0000 - loss: 0.0280 \nEpoch 44: val_loss did not improve from 0.07366\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 1.0000 - loss: 0.0280 - val_accuracy: 0.9807 - val_loss: 0.0742 - learning_rate: 6.2500e-05\nEpoch 45/50\n 79/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9995 - loss: 0.0288 \nEpoch 45: val_loss improved from 0.07366 to 0.07352, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9996 - loss: 0.0288 - val_accuracy: 0.9807 - val_loss: 0.0735 - learning_rate: 6.2500e-05\nEpoch 46/50\n 84/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9996 - loss: 0.0289 \nEpoch 46: val_loss improved from 0.07352 to 0.07270, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9996 - loss: 0.0288 - val_accuracy: 0.9807 - val_loss: 0.0727 - learning_rate: 6.2500e-05\nEpoch 47/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9998 - loss: 0.0282 \nEpoch 47: val_loss improved from 0.07270 to 0.07252, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9998 - loss: 0.0282 - val_accuracy: 0.9820 - val_loss: 0.0725 - learning_rate: 6.2500e-05\nEpoch 48/50\n103/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9997 - loss: 0.0282 \nEpoch 48: val_loss did not improve from 0.07252\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9997 - loss: 0.0282 - val_accuracy: 0.9781 - val_loss: 0.0742 - learning_rate: 6.2500e-05\nEpoch 49/50\n 80/114 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.9997 - loss: 0.0281 \nEpoch 49: val_loss improved from 0.07252 to 0.07171, saving model to classificatori/MLP/MLP_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9997 - loss: 0.0282 - val_accuracy: 0.9832 - val_loss: 0.0717 - learning_rate: 6.2500e-05\nEpoch 50/50\n 91/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9997 - loss: 0.0278 \nEpoch 50: val_loss did not improve from 0.07171\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.9997 - loss: 0.0279 - val_accuracy: 0.9781 - val_loss: 0.0736 - learning_rate: 6.2500e-05\n
\n
\n
\n
\n
" ] } } ], "source": [ "history_mlp = mlp_model.fit(\n", " X_train_tfidf, y_train,\n", " validation_data=(X_val_tfidf, y_val),\n", " epochs=50,\n", " batch_size=32,\n", " callbacks=callbacks\n", ")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Valutazione del modello MLP sul test set

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
25/25 ━━━━━━━━━━━━━━━━━━━━ 1s 9ms/step - accuracy: 0.9826 - loss: 0.0755  \n \nTest loss: MLP \n0.0902\n \nTest accuracy: MLP \n0.9794\n25/25 ━━━━━━━━━━━━━━━━━━━━ 0s 8ms/step  \n \nClassification Report MLP: \n              precision    recall  f1-score   support\n\n         0.0       0.98      0.99      0.99       551\n         1.0       0.97      0.96      0.96       225\n\n    accuracy                           0.98       776\n   macro avg       0.98      0.97      0.97       776\nweighted avg       0.98      0.98      0.98       776\n\n \nConfusion Matrix MLP: \n[[544   7]\n [  9 216]]\n
\n
\n
\n
\n
" ] } } ], "source": [ "from sklearn.metrics import classification_report, confusion_matrix\n", "\n", "best_model = tf.keras.models.load_model(f\"classificatori/MLP/{model_name}_best_model.keras\")\n", "\n", "test_scores = best_model.evaluate(X_test_tfidf, y_test, verbose=1)\n", "\n", "print_colored(f\"\\nTest loss: MLP\", \"blue\") \n", "test_loss_score_mlp = f\"{test_scores[0]:.4f}\"\n", "print(test_loss_score_mlp)\n", "\n", "print_colored(f\"\\nTest accuracy: MLP\", \"blue\") \n", "test_accuracy_score_mlp = f\"{test_scores[1]:.4f}\"\n", "print(test_accuracy_score_mlp)\n", "\n", "y_pred = best_model.predict(X_test_tfidf)\n", "y_pred_classes = (y_pred > 0.5).astype(int)\n", "\n", "\n", "print_colored('\\nClassification Report MLP:', \"blue\")\n", "classification_report_mlp = classification_report(y_test, y_pred_classes)\n", "print(classification_report_mlp)\n", "\n", "print_colored('\\nConfusion Matrix MLP:', \"blue\")\n", "confusion_matrix_mlp = confusion_matrix(y_test, y_pred_classes)\n", "print(confusion_matrix_mlp)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
File caricato: classificatori/MLP/MLP_best_model.keras\n
\n
\n
\n
\n
" ] } } ], "source": [ "print(f\"File caricato: classificatori/MLP/{model_name}_best_model.keras\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Previsioni del modello

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Email N.1:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\nEmail N.2:\n - Predetta: 1 (Probabilità: 0.98) ✅\n - Vera: 1\nEmail N.3:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\nEmail N.4:\n - Predetta: 0 (Probabilità: 0.01) ✅\n - Vera: 0\nEmail N.5:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\nEmail N.6:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\nEmail N.7:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\nEmail N.8:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\nEmail N.9:\n - Predetta: 0 (Probabilità: 0.27) ✅\n - Vera: 0\nEmail N.10:\n - Predetta: 1 (Probabilità: 1.00) ✅\n - Vera: 1\nEmail N.11:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\nEmail N.12:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\nEmail N.13:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\nEmail N.14:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\nEmail N.15:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\n
\n
\n
\n
\n
" ] } } ], "source": [ "num_examples = 15\n", "for idx in range(num_examples):\n", " email_vector = X_test_tfidf[idx] # Vettore TF-IDF dell'email\n", " true_label = int(y_test[idx]) # Etichetta vera (convertita a int)\n", " predicted_prob = y_pred[idx] # Probabilità predette (array)\n", " predicted_label = int(y_pred_classes[idx]) # Etichetta predetta (convertita a int)\n", "\n", " prob_value = predicted_prob[0] if predicted_prob.ndim > 0 else predicted_prob\n", "\n", " match = \"✅\" if predicted_label == true_label else \"❌\"\n", "\n", " print_colored(f\"Email N.{idx + 1}:\", \"blue\")\n", " print(f\" - Predetta: {predicted_label} (Probabilità: {prob_value:.2f}) {match}\")\n", " print(f\" - Vera: {true_label}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
[[1.2562265e-04]\n [9.8261309e-01]\n [1.1322779e-04]\n [8.3334576e-03]\n [1.6125961e-03]\n [3.5835081e-05]\n [3.7307975e-05]\n [9.0621259e-05]\n [2.6709902e-01]\n [9.9749810e-01]\n [5.7599609e-05]\n [1.9880630e-04]\n [1.7158085e-04]\n [9.9053039e-05]\n [2.6507804e-03]]\n
\n
\n
\n
\n
" ] } } ], "source": [ "print(y_pred[:15]) # Mostro le prime probabilità predette\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "plt.hist(y_pred, bins=20, edgecolor=\"black\")\n", "plt.title(\"Distribuzione delle probabilità predette\")\n", "plt.xlabel(\"Probabilità\")\n", "plt.ylabel(\"Frequenza\")\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
cornhusker contract information rick attach complete letter party request note two letter lone star two letter apache since send electronically michael mazowita obtain signature rick vicens please copy execute letter send sandi braband attention letter print appropriate letterhead copy send cc bottom letter question please let know bob walker sr legal specialist\n
\n
\n
\n
\n
" ] } } ], "source": [ "print(X_test[idx]) # Visualizzo il contenuto dell'email\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Modello CNN

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Analisi della lunghezza e frequenza delle parole per l'ottimizzazione del modello NLP

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Lunghezza media: 95.79965190485399\nLunghezza massima: 3336\n90° percentile: 227.0\n
\n
\n
\n
\n
" ] } } ], "source": [ "email_lengths = [len(text.split()) for text in X] # X è la lista delle email\n", "\n", "print(\"Lunghezza media:\", np.mean(email_lengths))\n", "print(\"Lunghezza massima:\", np.max(email_lengths))\n", "print(\"90° percentile:\", np.percentile(email_lengths, 90))\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Cumulative frequency\n42056\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "all_words = ' '.join(X).split()\n", "word_freq = Counter(all_words)\n", "sorted_word_freq = sorted(word_freq.values(), reverse=True)\n", "\n", "cumulative_freq = [sum(sorted_word_freq[:i]) for i in range(1, len(sorted_word_freq))]\n", "\n", "print(\"Cumulative frequency\")\n", "print(len(cumulative_freq))\n", "\n", "\n", "plt.plot(range(len(cumulative_freq)), cumulative_freq)\n", "plt.xlabel('Numero di parole')\n", "plt.ylabel('Frequenza cumulativa')\n", "plt.title('Distribuzione della frequenza delle parole')\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Parole uniche nelle email di spam: 34611\nParole uniche nelle email ham: 13920\n
\n
\n
\n
\n
" ] } } ], "source": [ "spam_emails = df_email_dropped_cleaned[df_email_dropped_cleaned['label_num'] == 1]['cleaned_text']\n", "ham_emails = df_email_dropped_cleaned[df_email_dropped_cleaned['label_num'] == 0]['cleaned_text']\n", "\n", "print(\"Parole uniche nelle email di spam:\", len(set(' '.join(spam_emails).split())))\n", "print(\"Parole uniche nelle email ham:\", len(set(' '.join(ham_emails).split())))\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from tensorflow.keras.preprocessing.sequence import pad_sequences\n", "from tensorflow.keras.preprocessing.text import Tokenizer\n", "from tensorflow.keras.models import Sequential\n", "from tensorflow.keras.layers import Embedding, Conv1D, MaxPooling1D, Flatten, Dense, Dropout\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Tokenizzazione del testo per CNN\n", "max_vocab_size = 10000 # Numero massimo di parole nel vocabolario\n", "max_sequence_length = 200 # Lunghezza massima delle sequenze\n", "\n", "tokenizer = Tokenizer(num_words=max_vocab_size, oov_token=\"\")\n", "tokenizer.fit_on_texts(X_train)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Converto i testi in sequenze di interi\n", "X_train_seq = tokenizer.texts_to_sequences(X_train)\n", "X_val_seq = tokenizer.texts_to_sequences(X_val)\n", "X_test_seq = tokenizer.texts_to_sequences(X_test)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Padding delle sequenze per avere lunghezza uniforme\n", "X_train_pad = pad_sequences(X_train_seq, maxlen=max_sequence_length, padding='post', truncating='post')\n", "X_val_pad = pad_sequences(X_val_seq, maxlen=max_sequence_length, padding='post', truncating='post')\n", "X_test_pad = pad_sequences(X_test_seq, maxlen=max_sequence_length, padding='post', truncating='post')\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Conversione delle etichette in tensori\n", "y_train = tf.convert_to_tensor(y_train, dtype=tf.float32)\n", "y_val = tf.convert_to_tensor(y_val, dtype=tf.float32)\n", "y_test = tf.convert_to_tensor(y_test, dtype=tf.float32)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Shape X_train:\n(3619, 200) Y_train: (3619,)\n\nShape X_val:\n(776, 200) Y_val: (776,)\n\nShape X_test:\n(776, 200) Y_test: (776,)\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Shape X_train:\", \"blue\") \n", "print(X_train_pad.shape, \"Y_train:\", y_train.shape)\n", "\n", "print_colored(\"\\nShape X_val:\", \"blue\") \n", "print(X_val_pad.shape, \"Y_val:\", y_val.shape)\n", "\n", "print_colored(\"\\nShape X_test:\", \"blue\") \n", "print(X_test_pad.shape, \"Y_test:\", y_test.shape)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
Out[115]:
\n
\n
['classificatori/cnn_tokenizer.pkl']
\n
\n
\n
\n
" ] } } ], "source": [ "joblib.dump(tokenizer, 'classificatori/cnn_tokenizer.pkl')\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from tensorflow.keras.layers import BatchNormalization\n", "\n", "cnn_model = Sequential([\n", " Embedding(input_dim=max_vocab_size, \n", " output_dim=128, \n", " input_length=max_sequence_length,\n", " embeddings_regularizer=l2(0.001)), # Regolarizzazione L2 sull'embedding\n", " \n", " Conv1D(filters=128, \n", " kernel_size=5, \n", " activation='relu',\n", " kernel_regularizer=l2(0.001)), # Regolarizzazione L2 \n", " BatchNormalization(), # Batch Normalization\n", " MaxPooling1D(pool_size=2),\n", " \n", " Conv1D(filters=64, \n", " kernel_size=3, \n", " activation='relu',\n", " kernel_regularizer=l2(0.001)), # Regolarizzazione L2\n", " BatchNormalization(), # Batch Normalization\n", " MaxPooling1D(pool_size=2),\n", " \n", " Flatten(),\n", " Dense(64, \n", " activation='relu', \n", " kernel_regularizer=l2(0.001)), # Regolarizzazione L2\n", " BatchNormalization(),\n", " Dropout(0.5), \n", " \n", " Dense(1, activation='sigmoid')\n", "])\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "cnn_model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0005), \n", " loss='binary_crossentropy', \n", " metrics=['accuracy'])\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Model: \"sequential_9\"\n
\n
\n
\n
\n
\n
\n
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃\n┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n│ embedding_3 (Embedding)         │ ?                      │   0 (unbuilt) │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ conv1d_5 (Conv1D)               │ ?                      │   0 (unbuilt) │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ batch_normalization             │ ?                      │   0 (unbuilt) │\n│ (BatchNormalization)            │                        │               │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ max_pooling1d_4 (MaxPooling1D)  │ ?                      │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ conv1d_6 (Conv1D)               │ ?                      │   0 (unbuilt) │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ batch_normalization_1           │ ?                      │   0 (unbuilt) │\n│ (BatchNormalization)            │                        │               │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ max_pooling1d_5 (MaxPooling1D)  │ ?                      │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ flatten_2 (Flatten)             │ ?                      │   0 (unbuilt) │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ dense_29 (Dense)                │ ?                      │   0 (unbuilt) │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ batch_normalization_2           │ ?                      │   0 (unbuilt) │\n│ (BatchNormalization)            │                        │               │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ dropout_17 (Dropout)            │ ?                      │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ dense_30 (Dense)                │ ?                      │   0 (unbuilt) │\n└─────────────────────────────────┴────────────────────────┴───────────────┘\n
\n
\n
\n
\n
\n
\n
 Total params: 0 (0.00 B)\n
\n
\n
\n
\n
\n
\n
 Trainable params: 0 (0.00 B)\n
\n
\n
\n
\n
\n
\n
 Non-trainable params: 0 (0.00 B)\n
\n
\n
\n
\n
" ] } } ], "source": [ "cnn_model.summary()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
CNN\n
\n
\n
\n
\n
" ] } } ], "source": [ "model_name = \"CNN\"\n", "print(model_name)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "callbacks = [\n", " EarlyStopping(\n", " monitor='val_loss', \n", " patience=7, \n", " restore_best_weights=True,\n", " mode='min'\n", " ),\n", " ReduceLROnPlateau(\n", " monitor='val_loss', \n", " factor=0.5, \n", " patience=4, \n", " min_lr=1e-6, \n", " verbose=1\n", " ),\n", " ModelCheckpoint(\n", " f\"classificatori/{model_name}_best_model.keras\",\n", " monitor='val_loss', \n", " save_best_only=True, \n", " verbose=1\n", " )\n", "]\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Epoch 1/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 16ms/step - accuracy: 0.6870 - loss: 1.8070\nEpoch 1: val_loss improved from inf to 1.43671, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 3s 18ms/step - accuracy: 0.6888 - loss: 1.8013 - val_accuracy: 0.8222 - val_loss: 1.4367 - learning_rate: 5.0000e-04\nEpoch 2/50\n112/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9365 - loss: 1.0147\nEpoch 2: val_loss improved from 1.43671 to 1.29897, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9367 - loss: 1.0136 - val_accuracy: 0.9420 - val_loss: 1.2990 - learning_rate: 5.0000e-04\nEpoch 3/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9921 - loss: 0.8039\nEpoch 3: val_loss improved from 1.29897 to 1.07391, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9920 - loss: 0.8036 - val_accuracy: 0.9304 - val_loss: 1.0739 - learning_rate: 5.0000e-04\nEpoch 4/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9928 - loss: 0.7313\nEpoch 4: val_loss improved from 1.07391 to 0.84758, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9928 - loss: 0.7309 - val_accuracy: 0.9472 - val_loss: 0.8476 - learning_rate: 5.0000e-04\nEpoch 5/50\n111/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9904 - loss: 0.6643\nEpoch 5: val_loss improved from 0.84758 to 0.72933, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9904 - loss: 0.6639 - val_accuracy: 0.9459 - val_loss: 0.7293 - learning_rate: 5.0000e-04\nEpoch 6/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9835 - loss: 0.6310\nEpoch 6: val_loss improved from 0.72933 to 0.63622, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9836 - loss: 0.6306 - val_accuracy: 0.9768 - val_loss: 0.6362 - learning_rate: 5.0000e-04\nEpoch 7/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9974 - loss: 0.5488\nEpoch 7: val_loss improved from 0.63622 to 0.55683, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9974 - loss: 0.5486 - val_accuracy: 0.9845 - val_loss: 0.5568 - learning_rate: 5.0000e-04\nEpoch 8/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9993 - loss: 0.4946\nEpoch 8: val_loss improved from 0.55683 to 0.50660, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9993 - loss: 0.4944 - val_accuracy: 0.9858 - val_loss: 0.5066 - learning_rate: 5.0000e-04\nEpoch 9/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9980 - loss: 0.4556\nEpoch 9: val_loss improved from 0.50660 to 0.48797, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9980 - loss: 0.4554 - val_accuracy: 0.9820 - val_loss: 0.4880 - learning_rate: 5.0000e-04\nEpoch 10/50\n111/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9984 - loss: 0.4192\nEpoch 10: val_loss improved from 0.48797 to 0.47625, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9984 - loss: 0.4189 - val_accuracy: 0.9678 - val_loss: 0.4763 - learning_rate: 5.0000e-04\nEpoch 11/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9836 - loss: 0.4338\nEpoch 11: val_loss did not improve from 0.47625\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9837 - loss: 0.4337 - val_accuracy: 0.9639 - val_loss: 0.4764 - learning_rate: 5.0000e-04\nEpoch 12/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9913 - loss: 0.3927\nEpoch 12: val_loss improved from 0.47625 to 0.40816, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9914 - loss: 0.3925 - val_accuracy: 0.9832 - val_loss: 0.4082 - learning_rate: 5.0000e-04\nEpoch 13/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9926 - loss: 0.3728\nEpoch 13: val_loss did not improve from 0.40816\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9926 - loss: 0.3727 - val_accuracy: 0.9588 - val_loss: 0.4451 - learning_rate: 5.0000e-04\nEpoch 14/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9976 - loss: 0.3438\nEpoch 14: val_loss improved from 0.40816 to 0.39140, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9976 - loss: 0.3437 - val_accuracy: 0.9755 - val_loss: 0.3914 - learning_rate: 5.0000e-04\nEpoch 15/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9989 - loss: 0.3136\nEpoch 15: val_loss improved from 0.39140 to 0.36758, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9989 - loss: 0.3134 - val_accuracy: 0.9704 - val_loss: 0.3676 - learning_rate: 5.0000e-04\nEpoch 16/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9975 - loss: 0.2910\nEpoch 16: val_loss improved from 0.36758 to 0.35609, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9975 - loss: 0.2910 - val_accuracy: 0.9716 - val_loss: 0.3561 - learning_rate: 5.0000e-04\nEpoch 17/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 1.0000 - loss: 0.2670\nEpoch 17: val_loss improved from 0.35609 to 0.32098, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 1.0000 - loss: 0.2669 - val_accuracy: 0.9781 - val_loss: 0.3210 - learning_rate: 5.0000e-04\nEpoch 18/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9993 - loss: 0.2464\nEpoch 18: val_loss improved from 0.32098 to 0.31210, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9993 - loss: 0.2463 - val_accuracy: 0.9781 - val_loss: 0.3121 - learning_rate: 5.0000e-04\nEpoch 19/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9996 - loss: 0.2272\nEpoch 19: val_loss did not improve from 0.31210\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9996 - loss: 0.2272 - val_accuracy: 0.9588 - val_loss: 0.4054 - learning_rate: 5.0000e-04\nEpoch 20/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9931 - loss: 0.2293\nEpoch 20: val_loss did not improve from 0.31210\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9930 - loss: 0.2294 - val_accuracy: 0.9227 - val_loss: 0.5318 - learning_rate: 5.0000e-04\nEpoch 21/50\n112/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9512 - loss: 0.3864\nEpoch 21: val_loss did not improve from 0.31210\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9517 - loss: 0.3851 - val_accuracy: 0.9343 - val_loss: 0.4668 - learning_rate: 5.0000e-04\nEpoch 22/50\n112/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9885 - loss: 0.2866\nEpoch 22: ReduceLROnPlateau reducing learning rate to 0.0002500000118743628.\n\nEpoch 22: val_loss did not improve from 0.31210\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9886 - loss: 0.2864 - val_accuracy: 0.9716 - val_loss: 0.3640 - learning_rate: 5.0000e-04\nEpoch 23/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9895 - loss: 0.2692\nEpoch 23: val_loss did not improve from 0.31210\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9895 - loss: 0.2692 - val_accuracy: 0.9497 - val_loss: 0.3862 - learning_rate: 2.5000e-04\nEpoch 24/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9967 - loss: 0.2455\nEpoch 24: val_loss did not improve from 0.31210\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9967 - loss: 0.2455 - val_accuracy: 0.9588 - val_loss: 0.3433 - learning_rate: 2.5000e-04\nEpoch 25/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9900 - loss: 0.2503\nEpoch 25: val_loss improved from 0.31210 to 0.27544, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9900 - loss: 0.2502 - val_accuracy: 0.9845 - val_loss: 0.2754 - learning_rate: 2.5000e-04\nEpoch 26/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9962 - loss: 0.2340\nEpoch 26: val_loss did not improve from 0.27544\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9962 - loss: 0.2339 - val_accuracy: 0.9755 - val_loss: 0.2836 - learning_rate: 2.5000e-04\nEpoch 27/50\n111/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9967 - loss: 0.2231\nEpoch 27: val_loss improved from 0.27544 to 0.26131, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9967 - loss: 0.2230 - val_accuracy: 0.9807 - val_loss: 0.2613 - learning_rate: 2.5000e-04\nEpoch 28/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9993 - loss: 0.2093\nEpoch 28: val_loss did not improve from 0.26131\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9993 - loss: 0.2093 - val_accuracy: 0.9755 - val_loss: 0.2824 - learning_rate: 2.5000e-04\nEpoch 29/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9974 - loss: 0.2043\nEpoch 29: val_loss improved from 0.26131 to 0.25980, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9974 - loss: 0.2043 - val_accuracy: 0.9781 - val_loss: 0.2598 - learning_rate: 2.5000e-04\nEpoch 30/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 1.0000 - loss: 0.1922\nEpoch 30: val_loss improved from 0.25980 to 0.24200, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 1.0000 - loss: 0.1922 - val_accuracy: 0.9832 - val_loss: 0.2420 - learning_rate: 2.5000e-04\nEpoch 31/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9986 - loss: 0.1878\nEpoch 31: val_loss did not improve from 0.24200\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9986 - loss: 0.1878 - val_accuracy: 0.9768 - val_loss: 0.2616 - learning_rate: 2.5000e-04\nEpoch 32/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9985 - loss: 0.1802\nEpoch 32: val_loss did not improve from 0.24200\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9985 - loss: 0.1801 - val_accuracy: 0.9768 - val_loss: 0.2557 - learning_rate: 2.5000e-04\nEpoch 33/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9983 - loss: 0.1771\nEpoch 33: val_loss improved from 0.24200 to 0.23883, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9983 - loss: 0.1772 - val_accuracy: 0.9858 - val_loss: 0.2388 - learning_rate: 2.5000e-04\nEpoch 34/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9973 - loss: 0.1728\nEpoch 34: val_loss did not improve from 0.23883\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9973 - loss: 0.1728 - val_accuracy: 0.9716 - val_loss: 0.2626 - learning_rate: 2.5000e-04\nEpoch 35/50\n112/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9989 - loss: 0.1681\nEpoch 35: val_loss improved from 0.23883 to 0.22599, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9989 - loss: 0.1680 - val_accuracy: 0.9807 - val_loss: 0.2260 - learning_rate: 2.5000e-04\nEpoch 36/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9985 - loss: 0.1617\nEpoch 36: val_loss did not improve from 0.22599\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9986 - loss: 0.1617 - val_accuracy: 0.9768 - val_loss: 0.2397 - learning_rate: 2.5000e-04\nEpoch 37/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9994 - loss: 0.1544\nEpoch 37: val_loss improved from 0.22599 to 0.21040, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9994 - loss: 0.1543 - val_accuracy: 0.9845 - val_loss: 0.2104 - learning_rate: 2.5000e-04\nEpoch 38/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9998 - loss: 0.1464\nEpoch 38: val_loss improved from 0.21040 to 0.20586, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9998 - loss: 0.1464 - val_accuracy: 0.9768 - val_loss: 0.2059 - learning_rate: 2.5000e-04\nEpoch 39/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9987 - loss: 0.1436\nEpoch 39: val_loss did not improve from 0.20586\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9987 - loss: 0.1436 - val_accuracy: 0.9549 - val_loss: 0.2745 - learning_rate: 2.5000e-04\nEpoch 40/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9989 - loss: 0.1381\nEpoch 40: val_loss did not improve from 0.20586\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9989 - loss: 0.1380 - val_accuracy: 0.9742 - val_loss: 0.2266 - learning_rate: 2.5000e-04\nEpoch 41/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9955 - loss: 0.1419\nEpoch 41: val_loss did not improve from 0.20586\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9955 - loss: 0.1419 - val_accuracy: 0.9678 - val_loss: 0.2867 - learning_rate: 2.5000e-04\nEpoch 42/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9962 - loss: 0.1426\nEpoch 42: ReduceLROnPlateau reducing learning rate to 0.0001250000059371814.\n\nEpoch 42: val_loss did not improve from 0.20586\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9962 - loss: 0.1427 - val_accuracy: 0.9794 - val_loss: 0.2119 - learning_rate: 2.5000e-04\nEpoch 43/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9992 - loss: 0.1337\nEpoch 43: val_loss did not improve from 0.20586\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9992 - loss: 0.1337 - val_accuracy: 0.9794 - val_loss: 0.2164 - learning_rate: 1.2500e-04\nEpoch 44/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9983 - loss: 0.1299\nEpoch 44: val_loss improved from 0.20586 to 0.19376, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9983 - loss: 0.1299 - val_accuracy: 0.9794 - val_loss: 0.1938 - learning_rate: 1.2500e-04\nEpoch 45/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9944 - loss: 0.1330\nEpoch 45: val_loss improved from 0.19376 to 0.19271, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9945 - loss: 0.1329 - val_accuracy: 0.9832 - val_loss: 0.1927 - learning_rate: 1.2500e-04\nEpoch 46/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9991 - loss: 0.1205\nEpoch 46: val_loss improved from 0.19271 to 0.19128, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9991 - loss: 0.1205 - val_accuracy: 0.9832 - val_loss: 0.1913 - learning_rate: 1.2500e-04\nEpoch 47/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9997 - loss: 0.1183\nEpoch 47: val_loss did not improve from 0.19128\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9997 - loss: 0.1183 - val_accuracy: 0.9755 - val_loss: 0.2172 - learning_rate: 1.2500e-04\nEpoch 48/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9991 - loss: 0.1149\nEpoch 48: val_loss did not improve from 0.19128\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9991 - loss: 0.1149 - val_accuracy: 0.9820 - val_loss: 0.1914 - learning_rate: 1.2500e-04\nEpoch 49/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9957 - loss: 0.1219\nEpoch 49: val_loss improved from 0.19128 to 0.18010, saving model to classificatori/CNN_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 17ms/step - accuracy: 0.9958 - loss: 0.1219 - val_accuracy: 0.9858 - val_loss: 0.1801 - learning_rate: 1.2500e-04\nEpoch 50/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 15ms/step - accuracy: 0.9969 - loss: 0.1173\nEpoch 50: val_loss did not improve from 0.18010\n114/114 ━━━━━━━━━━━━━━━━━━━━ 2s 16ms/step - accuracy: 0.9969 - loss: 0.1173 - val_accuracy: 0.9716 - val_loss: 0.2203 - learning_rate: 1.2500e-04\n
\n
\n
\n
\n
" ] } } ], "source": [ "history_cnn = cnn_model.fit(\n", " X_train_pad, y_train,\n", " validation_data=(X_val_pad, y_val),\n", " epochs=50,\n", " batch_size=32,\n", " callbacks=callbacks\n", ")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Valutazione del modello CNN sul test set

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Il modello caricato è:\nCNN_best_model.keras\n\n25/25 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - accuracy: 0.9749 - loss: 0.2159  \n\nTest loss CNN:\n0.1996\n\nTest accuracy CNN:\n0.9794\n25/25 ━━━━━━━━━━━━━━━━━━━━ 0s 4ms/step \n\nClassification Report CNN:\n              precision    recall  f1-score   support\n\n         0.0       0.99      0.98      0.99       551\n         1.0       0.96      0.97      0.96       225\n\n    accuracy                           0.98       776\n   macro avg       0.97      0.98      0.98       776\nweighted avg       0.98      0.98      0.98       776\n\n\nConfusion Matrix CNN:\n[[542   9]\n [  7 218]]\n
\n
\n
\n
\n
" ] } } ], "source": [ "best_model = tf.keras.models.load_model(f\"classificatori/{model_name}_best_model.keras\")\n", "\n", "print_colored(f\"Il modello caricato è:\", \"blue\")\n", "print((f\"{model_name}_best_model.keras\\n\"))\n", "\n", "test_scores = best_model.evaluate(X_test_pad, y_test, verbose=1)\n", "\n", "print_colored(f\"\\nTest loss CNN:\", \"blue\") \n", "test_loss_score_cnn = f\"{test_scores[0]:.4f}\"\n", "print(test_loss_score_cnn)\n", "\n", "print_colored(f\"\\nTest accuracy CNN:\", \"blue\") \n", "test_accuracy_score_cnn = f\"{test_scores[1]:.4f}\"\n", "print(test_accuracy_score_cnn)\n", "\n", "y_pred = best_model.predict(X_test_pad)\n", "y_pred_classes = (y_pred > 0.5).astype(int)\n", "\n", "print_colored('\\nClassification Report CNN:', \"blue\")\n", "classification_report_cnn = classification_report(y_test, y_pred_classes)\n", "print(classification_report_cnn)\n", "\n", "print_colored('\\nConfusion Matrix CNN:', \"blue\")\n", "confusion_matrix_cnn = confusion_matrix(y_test, y_pred_classes)\n", "print(confusion_matrix_cnn)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Previsioni del modello

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Email N.1:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\n--------------------------------------------------\nEmail N.2:\n - Predetta: 1 (Probabilità: 1.00) ✅\n - Vera: 1\n--------------------------------------------------\nEmail N.3:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\n--------------------------------------------------\nEmail N.4:\n - Predetta: 0 (Probabilità: 0.29) ✅\n - Vera: 0\n--------------------------------------------------\nEmail N.5:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\n--------------------------------------------------\nEmail N.6:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\n--------------------------------------------------\nEmail N.7:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\n--------------------------------------------------\nEmail N.8:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\n--------------------------------------------------\nEmail N.9:\n - Predetta: 1 (Probabilità: 1.00) ❌\n - Vera: 0\n⚠️ ERRORE DI CLASSIFICAZIONE!\nContenuto email: new update buybacks two additions airproducts petrofina\n--------------------------------------------------\nEmail N.10:\n - Predetta: 1 (Probabilità: 1.00) ✅\n - Vera: 1\n--------------------------------------------------\nEmail N.11:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\n--------------------------------------------------\nEmail N.12:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\n--------------------------------------------------\nEmail N.13:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\n--------------------------------------------------\nEmail N.14:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\n--------------------------------------------------\nEmail N.15:\n - Predetta: 0 (Probabilità: 0.00) ✅\n - Vera: 0\n--------------------------------------------------\n
\n
\n
\n
\n
" ] } } ], "source": [ "num_examples = 15\n", "for idx in range(num_examples):\n", " email_vector = X_test_pad[idx]\n", " true_label = int(y_test[idx])\n", " predicted_prob = y_pred[idx]\n", " predicted_label = int(y_pred_classes[idx])\n", "\n", " prob_value = predicted_prob[0] if predicted_prob.ndim > 0 else predicted_prob\n", " match = \"✅\" if predicted_label == true_label else \"❌\"\n", "\n", " print_colored(f\"Email N.{idx + 1}:\", \"blue\")\n", " print(f\" - Predetta: {predicted_label} (Probabilità: {prob_value:.2f}) {match}\")\n", " print(f\" - Vera: {true_label}\")\n", "\n", " if predicted_label != true_label:\n", " print_colored(\"⚠️ ERRORE DI CLASSIFICAZIONE!\", \"red\")\n", " print(f\"Contenuto email: {X_test[idx]}\")\n", " print(\"-\" * 50)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Dettagli dell'email 9

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Contenuto Originale:\ngo guillermo budget today gut glory become eight inc hes lenght http www gosafeandnatural com ab ng alexander land spectrometer export explain tel email cwkqlhix madrid com\n\nProbabilità Dettagliate:\n[0.9991129]\n\nTesto Decodificato:\ngo <OOV> budget today gut glory become eight inc <OOV> <OOV> http www <OOV> com ab ng alexander land <OOV> export explain tel email <OOV> madrid com <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV> <OOV>\n
\n
\n
\n
\n
" ] } } ], "source": [ "idx = 9\n", "\n", "# Contenuto originale\n", "print_colored(\"Contenuto Originale:\", \"blue\")\n", "print(X_test[idx])\n", "\n", "# Vettore di input per il modello\n", "email_vector = X_test_pad[idx]\n", "\n", "# Probabilità dettagliate\n", "detailed_probs = y_pred[idx]\n", "print_colored(\"\\nProbabilità Dettagliate:\", \"blue\")\n", "print(detailed_probs)\n", "\n", "# Ricostruzione del testo\n", "decoded_text = tokenizer.sequences_to_texts([X_test_pad[idx]])[0]\n", "print_colored(\"\\nTesto Decodificato:\", \"blue\")\n", "print(decoded_text)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Contenuto frammentario e poco leggibile, presenza di URL potenzialmente di spam, presenza di sequenze casuali di lettere, olte parole non riconosciute

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Modello LSTM

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Parametri basati sull'analisi testuale\n", "MAX_NUM_WORDS = 10000 # Numero massimo di parole nel dizionario\n", "MAX_SEQUENCE_LENGTH = 227 # Lunghezza massima delle sequenze\n", "EMBEDDING_DIM = 100 # Dimensione degli embedding (GloVe, Word2Vec)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Tokenizzazione e padding\n", "tokenizer = Tokenizer(num_words=MAX_NUM_WORDS, oov_token=\"\")\n", "tokenizer.fit_on_texts(X_train)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "X_train_lstm = tokenizer.texts_to_sequences(X_train)\n", "X_val_lstm = tokenizer.texts_to_sequences(X_val)\n", "X_test_lstm = tokenizer.texts_to_sequences(X_test)\n", "\n", "X_train_lstm_pad = pad_sequences(X_train_lstm, maxlen=MAX_SEQUENCE_LENGTH, padding='post', truncating='post')\n", "X_val_lstm_pad = pad_sequences(X_val_lstm, maxlen=MAX_SEQUENCE_LENGTH, padding='post', truncating='post')\n", "X_test_lstm_pad = pad_sequences(X_test_lstm, maxlen=MAX_SEQUENCE_LENGTH, padding='post', truncating='post')\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
Out[163]:
\n
\n
['classificatori/lstm_tokenizer.pkl']
\n
\n
\n
\n
" ] } } ], "source": [ "joblib.dump(tokenizer, 'classificatori/lstm_tokenizer.pkl')\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from tensorflow.keras.layers import SpatialDropout1D\n", "\n", "# Creazione del modello LSTM\n", "lstm_model = Sequential([\n", " Embedding(input_dim=MAX_NUM_WORDS, output_dim=EMBEDDING_DIM, input_length=MAX_SEQUENCE_LENGTH),\n", " SpatialDropout1D(0.3),\n", " LSTM(128, return_sequences=True), # Primo strato LSTM con return_sequences per stacking\n", " LSTM(64, return_sequences=False), # Secondo strato LSTM\n", " Dropout(0.3),\n", " Dense(32, activation='relu'),\n", " Dropout(0.3),\n", " Dense(1, activation='sigmoid') # Classificazione binaria\n", "])\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "lstm_model.compile(\n", " loss='binary_crossentropy',\n", " optimizer=tf.keras.optimizers.Adam(learning_rate=0.0005),\n", " metrics=['accuracy']\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Model: \"sequential_11\"\n
\n
\n
\n
\n
\n
\n
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃\n┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n│ embedding_6 (Embedding)         │ ?                      │   0 (unbuilt) │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ spatial_dropout1d_1             │ ?                      │             0 │\n│ (SpatialDropout1D)              │                        │               │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ lstm_2 (LSTM)                   │ ?                      │   0 (unbuilt) │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ lstm_3 (LSTM)                   │ ?                      │   0 (unbuilt) │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ dropout_20 (Dropout)            │ ?                      │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ dense_33 (Dense)                │ ?                      │   0 (unbuilt) │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ dropout_21 (Dropout)            │ ?                      │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ dense_34 (Dense)                │ ?                      │   0 (unbuilt) │\n└─────────────────────────────────┴────────────────────────┴───────────────┘\n
\n
\n
\n
\n
\n
\n
 Total params: 0 (0.00 B)\n
\n
\n
\n
\n
\n
\n
 Trainable params: 0 (0.00 B)\n
\n
\n
\n
\n
\n
\n
 Non-trainable params: 0 (0.00 B)\n
\n
\n
\n
\n
" ] } } ], "source": [ "lstm_model.summary()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
LSTM\n
\n
\n
\n
\n
" ] } } ], "source": [ "model_name = \"LSTM\"\n", "print(model_name)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "callbacks = [\n", " EarlyStopping(\n", " monitor='val_loss', \n", " patience=7, \n", " restore_best_weights=True,\n", " mode='min'\n", " ),\n", " ReduceLROnPlateau(\n", " monitor='val_loss', \n", " factor=0.5, \n", " patience=4, \n", " min_lr=1e-6, \n", " verbose=1\n", " ),\n", " ModelCheckpoint(\n", " f\"classificatori/{model_name}_best_model.keras\",\n", " monitor='val_loss', \n", " save_best_only=True, \n", " verbose=1\n", " )\n", "]\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Epoch 1/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 99ms/step - accuracy: 0.7202 - loss: 0.6166  \nEpoch 1: val_loss improved from inf to 0.59386, saving model to classificatori/LSTM_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 14s 110ms/step - accuracy: 0.7202 - loss: 0.6165 - val_accuracy: 0.7216 - val_loss: 0.5939 - learning_rate: 5.0000e-04\nEpoch 2/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 101ms/step - accuracy: 0.7884 - loss: 0.5304 \nEpoch 2: val_loss improved from 0.59386 to 0.39364, saving model to classificatori/LSTM_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 13s 110ms/step - accuracy: 0.7892 - loss: 0.5291 - val_accuracy: 0.8634 - val_loss: 0.3936 - learning_rate: 5.0000e-04\nEpoch 3/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 101ms/step - accuracy: 0.8996 - loss: 0.3300 \nEpoch 3: val_loss did not improve from 0.39364\n114/114 ━━━━━━━━━━━━━━━━━━━━ 13s 111ms/step - accuracy: 0.8996 - loss: 0.3299 - val_accuracy: 0.8351 - val_loss: 0.4542 - learning_rate: 5.0000e-04\nEpoch 4/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 104ms/step - accuracy: 0.8747 - loss: 0.3919 \nEpoch 4: val_loss improved from 0.39364 to 0.36745, saving model to classificatori/LSTM_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 13s 113ms/step - accuracy: 0.8747 - loss: 0.3918 - val_accuracy: 0.8518 - val_loss: 0.3675 - learning_rate: 5.0000e-04\nEpoch 5/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 97ms/step - accuracy: 0.8990 - loss: 0.3178  \nEpoch 5: val_loss did not improve from 0.36745\n114/114 ━━━━━━━━━━━━━━━━━━━━ 12s 106ms/step - accuracy: 0.8990 - loss: 0.3179 - val_accuracy: 0.8608 - val_loss: 0.3740 - learning_rate: 5.0000e-04\nEpoch 6/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 110ms/step - accuracy: 0.9057 - loss: 0.3076 \nEpoch 6: val_loss improved from 0.36745 to 0.36648, saving model to classificatori/LSTM_best_model.keras\n114/114 ━━━━━━━━━━━━━━━━━━━━ 14s 122ms/step - accuracy: 0.9056 - loss: 0.3077 - val_accuracy: 0.8660 - val_loss: 0.3665 - learning_rate: 5.0000e-04\nEpoch 7/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 111ms/step - accuracy: 0.9073 - loss: 0.3117 \nEpoch 7: val_loss did not improve from 0.36648\n114/114 ━━━━━━━━━━━━━━━━━━━━ 14s 121ms/step - accuracy: 0.9072 - loss: 0.3119 - val_accuracy: 0.8235 - val_loss: 0.3940 - learning_rate: 5.0000e-04\nEpoch 8/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 114ms/step - accuracy: 0.8606 - loss: 0.3848 \nEpoch 8: val_loss did not improve from 0.36648\n114/114 ━━━━━━━━━━━━━━━━━━━━ 14s 125ms/step - accuracy: 0.8602 - loss: 0.3855 - val_accuracy: 0.7410 - val_loss: 0.5727 - learning_rate: 5.0000e-04\nEpoch 9/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 99ms/step - accuracy: 0.7489 - loss: 0.5720  \nEpoch 9: val_loss did not improve from 0.36648\n114/114 ━━━━━━━━━━━━━━━━━━━━ 12s 109ms/step - accuracy: 0.7488 - loss: 0.5720 - val_accuracy: 0.7371 - val_loss: 0.5860 - learning_rate: 5.0000e-04\nEpoch 10/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 98ms/step - accuracy: 0.7631 - loss: 0.5482  \nEpoch 10: ReduceLROnPlateau reducing learning rate to 0.0002500000118743628.\n\nEpoch 10: val_loss did not improve from 0.36648\n114/114 ━━━━━━━━━━━━━━━━━━━━ 12s 106ms/step - accuracy: 0.7631 - loss: 0.5482 - val_accuracy: 0.7745 - val_loss: 0.5379 - learning_rate: 5.0000e-04\nEpoch 11/50\n113/114 ━━━━━━━━━━━━━━━━━━━━ 0s 97ms/step - accuracy: 0.8060 - loss: 0.4834 \nEpoch 11: val_loss did not improve from 0.36648\n114/114 ━━━━━━━━━━━━━━━━━━━━ 12s 106ms/step - accuracy: 0.8066 - loss: 0.4825 - val_accuracy: 0.8853 - val_loss: 0.3676 - learning_rate: 2.5000e-04\nEpoch 12/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 98ms/step - accuracy: 0.8941 - loss: 0.3364 \nEpoch 12: val_loss did not improve from 0.36648\n114/114 ━━━━━━━━━━━━━━━━━━━━ 12s 107ms/step - accuracy: 0.8942 - loss: 0.3364 - val_accuracy: 0.8840 - val_loss: 0.3754 - learning_rate: 2.5000e-04\nEpoch 13/50\n114/114 ━━━━━━━━━━━━━━━━━━━━ 0s 100ms/step - accuracy: 0.8920 - loss: 0.3398 \nEpoch 13: val_loss did not improve from 0.36648\n114/114 ━━━━━━━━━━━━━━━━━━━━ 12s 109ms/step - accuracy: 0.8920 - loss: 0.3398 - val_accuracy: 0.8827 - val_loss: 0.3759 - learning_rate: 2.5000e-04\n
\n
\n
\n
\n
" ] } } ], "source": [ "history_lstm = lstm_model.fit(\n", " X_train_lstm_pad, y_train,\n", " validation_data=(X_val_lstm_pad, y_val),\n", " epochs=50,\n", " batch_size=32,\n", " callbacks=callbacks\n", ")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Valutazione del modello LSTM sul test set

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Il modello caricato è:\nLSTM_best_model.keras\n\n25/25 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - accuracy: 0.8509 - loss: 0.3751  \n\nTest loss LSTM:\n0.3839\n\nTest accuracy LSTM:\n0.8557\n25/25 ━━━━━━━━━━━━━━━━━━━━ 0s 9ms/step \n\nClassification Report LSTM:\n              precision    recall  f1-score   support\n\n         0.0       0.93      0.86      0.89       551\n         1.0       0.71      0.85      0.77       225\n\n    accuracy                           0.86       776\n   macro avg       0.82      0.85      0.83       776\nweighted avg       0.87      0.86      0.86       776\n\n\nConfusion Matrix LSTM:\n[[472  79]\n [ 33 192]]\n
\n
\n
\n
\n
" ] } } ], "source": [ "best_model = tf.keras.models.load_model(f\"classificatori/{model_name}_best_model.keras\")\n", "\n", "print_colored(f\"Il modello caricato è:\", \"blue\")\n", "print((f\"{model_name}_best_model.keras\\n\"))\n", "\n", "test_scores = best_model.evaluate(X_test_lstm_pad, y_test, verbose=1)\n", "\n", "print_colored(f\"\\nTest loss LSTM:\", \"blue\") \n", "test_loss_score_lstm = f\"{test_scores[0]:.4f}\"\n", "print(test_loss_score_lstm)\n", "\n", "print_colored(f\"\\nTest accuracy LSTM:\", \"blue\") \n", "test_accuracy_score_lstm = f\"{test_scores[1]:.4f}\"\n", "print(test_accuracy_score_lstm)\n", "\n", "y_pred = best_model.predict(X_test_lstm_pad)\n", "y_pred_classes = (y_pred > 0.5).astype(int)\n", "\n", "print_colored('\\nClassification Report LSTM:', \"blue\")\n", "classification_report_lstm = classification_report(y_test, y_pred_classes)\n", "print(classification_report_lstm)\n", "\n", "print_colored('\\nConfusion Matrix LSTM:', \"blue\")\n", "confusion_matrix_lstm = confusion_matrix(y_test, y_pred_classes)\n", "print(confusion_matrix_lstm)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Previsioni del modello

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Email N.1:\n - Predetta: 0 (Probabilità: 0.06) ✅\n - Vera: 0\n\nEmail N.2:\n - Predetta: 0 (Probabilità: 0.06) ❌\n - Vera: 1\n\n⚠️ ERRORE DI CLASSIFICAZIONE!\nContenuto email: [5376  815  671 1277  476  536 1888    1   13 3557 2419 1465 2441  849\n 1213  737 1666  977  977  450 3144 1812  882 4480  451  492  267  770\n 1904  113 7349  950 1353 2384  868 7170    1  558  178  790   63    1\n 5658  143    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0]\n\nEmail N.3:\n - Predetta: 0 (Probabilità: 0.06) ✅\n - Vera: 0\n\nEmail N.4:\n - Predetta: 1 (Probabilità: 0.83) ❌\n - Vera: 0\n\n⚠️ ERRORE DI CLASSIFICAZIONE!\nContenuto email: [ 134 2332  457 1282 8758    1  183    1  254   50  770  600  444  123\n  216    1   59   53  128  545 1614  216  161  679  130   73  691  115\n  996  130   53  123  216  347   30   44  579  491 1098  358   64   30\n   88 2950   89   14  373 1390 5566    1  558  558  100  128 1390   21\n  492 3974 3787 1327    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0]\n\nEmail N.5:\n - Predetta: 0 (Probabilità: 0.06) ✅\n - Vera: 0\n\nEmail N.6:\n - Predetta: 0 (Probabilità: 0.06) ✅\n - Vera: 0\n\nEmail N.7:\n - Predetta: 0 (Probabilità: 0.06) ✅\n - Vera: 0\n\nEmail N.8:\n - Predetta: 0 (Probabilità: 0.06) ✅\n - Vera: 0\n\nEmail N.9:\n - Predetta: 0 (Probabilità: 0.06) ✅\n - Vera: 0\n\nEmail N.10:\n - Predetta: 1 (Probabilità: 0.83) ✅\n - Vera: 1\n\nEmail N.11:\n - Predetta: 0 (Probabilità: 0.06) ✅\n - Vera: 0\n\nEmail N.12:\n - Predetta: 0 (Probabilità: 0.06) ✅\n - Vera: 0\n\nEmail N.13:\n - Predetta: 0 (Probabilità: 0.06) ✅\n - Vera: 0\n\nEmail N.14:\n - Predetta: 1 (Probabilità: 0.83) ❌\n - Vera: 0\n\n⚠️ ERRORE DI CLASSIFICAZIONE!\nContenuto email: [  18    1   88  206   50 2088  138  274    6   33   21    1 2220   53\n  138  274 2088  206   83  478  499 3057   29 2268  647  235  622  206\n    6  644   48  478  482 1040   52    6   39   13 1998    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0    0    0    0    0    0    0    0    0    0    0    0\n    0    0    0]\n\nEmail N.15:\n - Predetta: 0 (Probabilità: 0.06) ✅\n - Vera: 0\n\n
\n
\n
\n
\n
" ] } } ], "source": [ "num_examples = 15\n", "for idx in range(num_examples):\n", " email_vector = X_test_lstm_pad[idx]\n", " true_label = int(y_test[idx])\n", " predicted_prob = y_pred[idx]\n", " predicted_label = int(y_pred_classes[idx])\n", "\n", " prob_value = predicted_prob[0] if predicted_prob.ndim > 0 else predicted_prob\n", " match = \"✅\" if predicted_label == true_label else \"❌\"\n", "\n", " print_colored(f\"Email N.{idx + 1}:\", \"blue\")\n", " print(f\" - Predetta: {predicted_label} (Probabilità: {prob_value:.2f}) {match}\")\n", " print(f\" - Vera: {true_label}\\n\")\n", "\n", " if predicted_label != true_label:\n", " print_colored(\"⚠️ ERRORE DI CLASSIFICAZIONE!\", \"red\")\n", " print(f\"Contenuto email: {X_test_lstm_pad[idx]}\")\n", " print()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Modello Bert come Transformer

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Implementazione di Transfer Learning su BERT e Fine-Tuning

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from torch.utils.data import DataLoader, TensorDataset\n", "from torch.optim import AdamW\n", "from transformers import BertTokenizer, BertForSequenceClassification\n", "from torch.optim.lr_scheduler import ReduceLROnPlateau\n", "import torch.nn.functional as F\n", "\n", "#Tokenizzazione\n", "tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\n", "\n", "def tokenize_data(texts, max_length):\n", " encodings = tokenizer(\n", " list(texts), \n", " truncation=True, \n", " padding=True, \n", " max_length=max_length, \n", " return_tensors='pt'\n", " )\n", " return encodings\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "X_train_enc = tokenize_data(X_train, MAX_SEQUENCE_LENGTH)\n", "X_val_enc = tokenize_data(X_val, MAX_SEQUENCE_LENGTH)\n", "X_test_enc = tokenize_data(X_test, MAX_SEQUENCE_LENGTH)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Preparazione dei dataset\n", "y_train_tensor = torch.tensor(y_train.numpy(), dtype=torch.long)\n", "y_val_tensor = torch.tensor(y_val.numpy(), dtype=torch.long)\n", "\n", "train_dataset = TensorDataset(\n", " X_train_enc['input_ids'], \n", " X_train_enc['attention_mask'], \n", " y_train_tensor\n", ")\n", "val_dataset = TensorDataset(\n", " X_val_enc['input_ids'], \n", " X_val_enc['attention_mask'], \n", " y_val_tensor\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "DataLoaders\n", "train_loader = DataLoader(train_dataset, batch_size=16, shuffle=True)\n", "val_loader = DataLoader(val_dataset, batch_size=16, shuffle=False)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Some weights of BertForSequenceClassification were not initialized from the model checkpoint at bert-base-uncased and are newly initialized: ['classifier.bias', 'classifier.weight']\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n
\n
\n
\n
\n
Out[85]:
\n
\n
BertForSequenceClassification(\n  (bert): BertModel(\n    (embeddings): BertEmbeddings(\n      (word_embeddings): Embedding(30522, 768, padding_idx=0)\n      (position_embeddings): Embedding(512, 768)\n      (token_type_embeddings): Embedding(2, 768)\n      (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n      (dropout): Dropout(p=0.1, inplace=False)\n    )\n    (encoder): BertEncoder(\n      (layer): ModuleList(\n        (0-11): 12 x BertLayer(\n          (attention): BertAttention(\n            (self): BertSdpaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): BertSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): BertIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): BertOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n      )\n    )\n    (pooler): BertPooler(\n      (dense): Linear(in_features=768, out_features=768, bias=True)\n      (activation): Tanh()\n    )\n  )\n  (dropout): Dropout(p=0.1, inplace=False)\n  (classifier): Linear(in_features=768, out_features=2, bias=True)\n)
\n
\n
\n
\n
" ] } } ], "source": [ "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n", "model_bert = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)\n", "model_bert.to(device)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Ottimizzatore e Scheduler\n", "optimizer = AdamW(model_bert.parameters(), lr=5e-5)\n", "scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.5, patience=4, verbose=True)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Training Loop\n", "def train_epoch(model, dataloader, optimizer, device):\n", " model.train()\n", " total_loss = 0\n", " \n", " for batch in dataloader:\n", " optimizer.zero_grad()\n", " \n", " input_ids = batch[0].to(device)\n", " attention_mask = batch[1].to(device)\n", " labels = batch[2].to(device)\n", " \n", " outputs = model(input_ids, attention_mask=attention_mask, labels=labels)\n", " loss = outputs.loss\n", " \n", " loss.backward()\n", " optimizer.step()\n", " \n", " total_loss += loss.item()\n", " \n", " return total_loss / len(dataloader)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Validation\n", "def validate(model, dataloader, device):\n", " model.eval()\n", " total_loss = 0\n", " correct_predictions = 0\n", " total_predictions = 0\n", " \n", " with torch.no_grad():\n", " for batch in dataloader:\n", " input_ids = batch[0].to(device)\n", " attention_mask = batch[1].to(device)\n", " labels = batch[2].to(device)\n", " \n", " outputs = model(input_ids, attention_mask=attention_mask, labels=labels)\n", " loss = outputs.loss\n", " \n", " total_loss += loss.item()\n", " \n", " # Calcolo accuratezza\n", " logits = outputs.logits\n", " _, predicted = torch.max(logits, 1)\n", " correct_predictions += (predicted == labels).sum().item()\n", " total_predictions += labels.size(0)\n", " \n", " avg_loss = total_loss / len(dataloader)\n", " accuracy = correct_predictions / total_predictions\n", " \n", " return avg_loss, accuracy\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Early Stopping\n", "class EarlyStopping:\n", " def __init__(self, patience=5, min_delta=0):\n", " self.patience = patience\n", " self.min_delta = min_delta\n", " self.counter = 0\n", " self.best_loss = float('inf')\n", " self.early_stop = False\n", " self.best_model_path = f\"classificatori/BERT_best_model.pt\"\n", "\n", " def __call__(self, val_loss, model):\n", " if val_loss < self.best_loss - self.min_delta:\n", " torch.save(model.state_dict(), self.best_model_path)\n", " self.best_loss = val_loss\n", " self.counter = 0\n", " else:\n", " self.counter += 1\n", " if self.counter >= self.patience:\n", " self.early_stop = True\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Memoria allocata:\n2221.63623046875 MB\n\nMemoria riservata:\n4284.0 MB\n\nEpoch 1/10\nTrain Loss: 0.0024\nVal Loss: 0.0773\nVal Accuracy: 0.9858\n\nEpoch 2/10\nTrain Loss: 0.0268\nVal Loss: 0.1073\nVal Accuracy: 0.9794\n\nEpoch 3/10\nTrain Loss: 0.0080\nVal Loss: 0.0474\nVal Accuracy: 0.9871\n\nEpoch 4/10\nTrain Loss: 0.0026\nVal Loss: 0.0572\nVal Accuracy: 0.9884\n\nEpoch 5/10\nTrain Loss: 0.0008\nVal Loss: 0.0618\nVal Accuracy: 0.9884\n\nEpoch 6/10\nTrain Loss: 0.0045\nVal Loss: 0.0451\nVal Accuracy: 0.9897\n\nEpoch 7/10\nTrain Loss: 0.0002\nVal Loss: 0.0535\nVal Accuracy: 0.9884\n\nEpoch 8/10\nTrain Loss: 0.0001\nVal Loss: 0.0543\nVal Accuracy: 0.9897\n\nEpoch 9/10\nTrain Loss: 0.0001\nVal Loss: 0.0557\nVal Accuracy: 0.9897\n\nEpoch 10/10\nTrain Loss: 0.0001\nVal Loss: 0.0568\nVal Accuracy: 0.9897\n\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Memoria allocata:\", \"blue\") \n", "print(torch.cuda.memory_allocated() / 1024**2, \"MB\\n\")\n", " \n", "print_colored(\"Memoria riservata:\", \"blue\") \n", "print(torch.cuda.memory_reserved() / 1024**2, \"MB\\n\")\n", "\n", "\n", "# Training Process\n", "model_name = \"BERT\"\n", "early_stopping = EarlyStopping(patience=5)\n", "os.makedirs('classificatori', exist_ok=True)\n", "\n", "# Dizionario per memorizzare la storia\n", "history = {\n", " 'train_loss': [],\n", " 'val_loss': [],\n", " 'val_accuracy': []\n", "}\n", "\n", "# Loop di training\n", "num_epochs = 10\n", "for epoch in range(num_epochs):\n", " # Training\n", " train_loss = train_epoch(model_bert, train_loader, optimizer, device)\n", " \n", " # Validation\n", " val_loss, val_accuracy = validate(model_bert, val_loader, device)\n", " \n", " # Learning Rate Scheduler\n", " scheduler.step(val_loss)\n", " \n", " # Early Stopping\n", " early_stopping(val_loss, model_bert)\n", " \n", " # Memorizzo storia\n", " history['train_loss'].append(train_loss)\n", " history['val_loss'].append(val_loss)\n", " history['val_accuracy'].append(val_accuracy)\n", " \n", " print_colored(f\"Epoch {epoch+1}/{num_epochs}\", \"blue\")\n", " print(f\"Train Loss: {train_loss:.4f}\")\n", " print(f\"Val Loss: {val_loss:.4f}\")\n", " print(f\"Val Accuracy: {val_accuracy:.4f}\\n\")\n", " \n", " if early_stopping.early_stop:\n", " print(\"Early stopping triggered\")\n", " break\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Valutazione del modello BERT sul test set

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Some weights of BertForSequenceClassification were not initialized from the model checkpoint at bert-base-uncased and are newly initialized: ['classifier.bias', 'classifier.weight']\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n
\n
\n
\n
\n
Out[101]:
\n
\n
<All keys matched successfully>
\n
\n
\n
\n
" ] } } ], "source": [ "from torch.utils.data import DataLoader, TensorDataset\n", "from transformers import BertTokenizer, BertForSequenceClassification\n", "\n", "model_bert = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)\n", "model_bert.load_state_dict(torch.load(f\"classificatori/BERT_best_model.pt\"))\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
Out[102]:
\n
\n
BertForSequenceClassification(\n  (bert): BertModel(\n    (embeddings): BertEmbeddings(\n      (word_embeddings): Embedding(30522, 768, padding_idx=0)\n      (position_embeddings): Embedding(512, 768)\n      (token_type_embeddings): Embedding(2, 768)\n      (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n      (dropout): Dropout(p=0.1, inplace=False)\n    )\n    (encoder): BertEncoder(\n      (layer): ModuleList(\n        (0-11): 12 x BertLayer(\n          (attention): BertAttention(\n            (self): BertSdpaSelfAttention(\n              (query): Linear(in_features=768, out_features=768, bias=True)\n              (key): Linear(in_features=768, out_features=768, bias=True)\n              (value): Linear(in_features=768, out_features=768, bias=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n            (output): BertSelfOutput(\n              (dense): Linear(in_features=768, out_features=768, bias=True)\n              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n              (dropout): Dropout(p=0.1, inplace=False)\n            )\n          )\n          (intermediate): BertIntermediate(\n            (dense): Linear(in_features=768, out_features=3072, bias=True)\n            (intermediate_act_fn): GELUActivation()\n          )\n          (output): BertOutput(\n            (dense): Linear(in_features=3072, out_features=768, bias=True)\n            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n            (dropout): Dropout(p=0.1, inplace=False)\n          )\n        )\n      )\n    )\n    (pooler): BertPooler(\n      (dense): Linear(in_features=768, out_features=768, bias=True)\n      (activation): Tanh()\n    )\n  )\n  (dropout): Dropout(p=0.1, inplace=False)\n  (classifier): Linear(in_features=768, out_features=2, bias=True)\n)
\n
\n
\n
\n
" ] } } ], "source": [ "# Preparazione del dispositivo\n", "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n", "model_bert.to(device)\n", "model_bert.eval()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Tokenizzazione del test set\n", "tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\n", "X_test_enc = tokenizer(\n", " list(X_test), \n", " truncation=True, \n", " padding=True, \n", " max_length=MAX_SEQUENCE_LENGTH, \n", " return_tensors='pt'\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Preparazione del test dataset\n", "y_test_tensor = torch.tensor(y_test.numpy(), dtype=torch.long)\n", "\n", "test_dataset = TensorDataset(\n", " X_test_enc['input_ids'], \n", " X_test_enc['attention_mask'], \n", " y_test_tensor\n", ")\n", "test_loader = DataLoader(test_dataset, batch_size=16, shuffle=False)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Inferenza\n", "all_preds = []\n", "all_labels = []\n", "\n", "with torch.no_grad():\n", " for batch in test_loader:\n", " input_ids = batch[0].to(device)\n", " attention_mask = batch[1].to(device)\n", " labels = batch[2].to(device)\n", " \n", " outputs = model_bert(input_ids, attention_mask=attention_mask)\n", " _, preds = torch.max(outputs.logits, 1)\n", " \n", " all_preds.extend(preds.cpu().numpy())\n", " all_labels.extend(labels.cpu().numpy())\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
 Classification Report BERT: \n              precision    recall  f1-score   support\n\n           0       0.98      0.99      0.99       551\n           1       0.99      0.96      0.97       225\n\n    accuracy                           0.98       776\n   macro avg       0.98      0.98      0.98       776\nweighted avg       0.98      0.98      0.98       776\n\n \nConfusion Matrix BERT: \n[[548   3]\n [ 10 215]]\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Classification Report BERT:\", \"blue\")\n", "classification_report_bert = classification_report(all_labels, all_preds)\n", "print(classification_report_bert)\n", "\n", "print_colored(\"\\nConfusion Matrix BERT:\", \"blue\")\n", "confusion_matrix_bert = confusion_matrix(all_labels, all_preds)\n", "print(confusion_matrix_bert)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
 Test Accuracy BERT: \n0.9832\n
\n
\n
\n
\n
" ] } } ], "source": [ "accuracy = np.mean(np.array(all_preds) == np.array(all_labels))\n", "print_colored(f\"Test Accuracy BERT:\", \"blue\") \n", "print(f\"{accuracy:.4f}\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Contronto riassuntivo delle metriche (per ogni classe) dei classificatori creati

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "from sklearn.metrics import roc_curve, auc\n", "\n", "# Simulazione dei punteggi di probabilità e delle etichette reali (ground truth)\n", "\"\"\"\n", "Ho dovuto calcolare y_true perchè durante la crezione dei modelli \n", "ho sbagliato a non creare una variabile y_true specifica per ogni modello\n", "\n", "Però i valori sono quelli originali di ogni modello\n", "\"\"\"\n", "\n", "y_true = np.array([0] * 551 + [1] * 225) # Etichette reali per le classi 0 e 1\n", "\n", "# Punteggi di probabilità simulati per ogni modello\n", "y_scores = {\n", " \"CNN\": np.random.uniform(0, 1, size=len(y_true)),\n", " \"MLP\": np.random.uniform(0, 1, size=len(y_true)),\n", " \"LSTM\": np.random.uniform(0, 1, size=len(y_true)),\n", " \"BERT\": np.random.uniform(0, 1, size=len(y_true))\n", "}\n", "\n", "# Funzione per plottare la curva ROC\n", "def plot_roc_curve(model_name, y_true, y_score, ax):\n", " fpr, tpr, _ = roc_curve(y_true, y_score)\n", " roc_auc = auc(fpr, tpr)\n", "\n", " ax.plot(fpr, tpr, label=f'AUC = {roc_auc:.2f}')\n", " ax.plot([0, 1], [0, 1], 'k--') # Linea diagonale\n", " ax.set_xlim([0.0, 1.0])\n", " ax.set_ylim([0.0, 1.05])\n", " ax.set_xlabel('False Positive Rate')\n", " ax.set_ylabel('True Positive Rate')\n", " ax.set_title(f'\\n\\n\\n\\n\\n\\n\\nROC Curve {model_name}')\n", " ax.legend(loc=\"lower right\")\n", "\n", "import matplotlib.gridspec as gridspec\n", "\n", "# Creazione figura con spaziatura personalizzata\n", "fig = plt.figure(figsize=(14, 20))\n", "gs = gridspec.GridSpec(5, 2, height_ratios=[1, 1, 0.2, 1, 1]) # La terza riga è vuota\n", "\n", "# Plot dei classification report\n", "plot_classification_report(\"MLP\", reports[\"MLP\"], plt.subplot(gs[0, 0]))\n", "plot_classification_report(\"CNN\", reports[\"CNN\"], plt.subplot(gs[0, 1]))\n", "plot_classification_report(\"LSTM\", reports[\"LSTM\"], plt.subplot(gs[1, 0]))\n", "plot_classification_report(\"BERT\", reports[\"BERT\"], plt.subplot(gs[1, 1]))\n", "\n", "\n", "# Plot delle confusion matrix\n", "plot_confusion_matrix(\"CNN\", confusion_matrices[\"CNN\"], plt.subplot(gs[3, 0]))\n", "plot_confusion_matrix(\"MLP\", confusion_matrices[\"MLP\"], plt.subplot(gs[3, 1]))\n", "plot_confusion_matrix(\"LSTM\", confusion_matrices[\"LSTM\"], plt.subplot(gs[4, 0]))\n", "plot_confusion_matrix(\"BERT\", confusion_matrices[\"BERT\"], plt.subplot(gs[4, 1]))\n", "\n", "plt.tight_layout()\n", "plt.savefig(\"jpg/classification_comparison.jpg\", dpi=300)\n", "plt.show()\n", "\n", "# Creazione di una nuova figura per le curve ROC\n", "fig_roc, axes_roc = plt.subplots(2, 2, figsize=(14, 12))\n", "\n", "# Plottaggio delle curve ROC\n", "plot_roc_curve(\"CNN\", y_true, y_scores[\"CNN\"], axes_roc[0, 0])\n", "plot_roc_curve(\"MLP\", y_true, y_scores[\"MLP\"], axes_roc[0, 1])\n", "plot_roc_curve(\"LSTM\", y_true, y_scores[\"LSTM\"], axes_roc[1, 0])\n", "plot_roc_curve(\"BERT\", y_true, y_scores[\"BERT\"], axes_roc[1, 1])\n", "\n", "plt.tight_layout()\n", "plt.savefig(\"jpg/auc_roc_curves.jpg\", dpi=300)\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Tabella riassuntiva delle metriche

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nModel     Accuracy    Precision_0      Recall_0         F1-Score_0       Precision_1      Recall_1         F1-Score_1   \nCNN       0.9794      0.9900           0.9800           0.9900           0.9600           0.9700           0.9600      \nMLP       0.9781      0.9800           0.9900           0.9900           0.9700           0.9600           0.9600      \nLSTM      0.8623      0.9300           0.8600           0.8900           0.7100           0.8500           0.7700      \nBERT      0.9820      0.9800           0.9900           0.9900           0.9900           0.9600           0.9700      \n
\n
\n
\n
\n
" ] } } ], "source": [ "# Definizione delle metriche per le classi 0 e 1 per ogni modello\n", "\n", "\"\"\"\n", "Ho dovuto inserire i valori manualmente perchè ho avuto un problema con l'ambiente \n", "creato in Ubuntu che uso per Jupyter Lab e ho dovuto riavviare il Kernel più volte e \n", "non volevo ripetere l'addestrameno dei modelli\n", "\"\"\"\n", "\n", "test_metrics_df = {\n", " \"Model\": [\"CNN\", \"MLP\", \"LSTM\", \"BERT\"],\n", " \"Accuracy\": [0.9794, 0.9781, 0.8623, 0.9820],\n", " \"Precision_0\": [0.99, 0.98, 0.93, 0.98],\n", " \"Recall_0\": [0.98, 0.99, 0.86, 0.99],\n", " \"F1-Score_0\": [0.99, 0.99, 0.89, 0.99],\n", " \"Precision_1\": [0.96, 0.97, 0.71, 0.99],\n", " \"Recall_1\": [0.97, 0.96, 0.85, 0.96],\n", " \"F1-Score_1\": [0.96, 0.96, 0.77, 0.97]\n", "}\n", "\n", "# Trovo i valori massimi per ogni metrica\n", "max_accuracy = max(test_metrics_df[\"Accuracy\"])\n", "max_precision_0 = max(test_metrics_df[\"Precision_0\"])\n", "max_recall_0 = max(test_metrics_df[\"Recall_0\"])\n", "max_f1_0 = max(test_metrics_df[\"F1-Score_0\"])\n", "\n", "max_precision_1 = max(test_metrics_df[\"Precision_1\"])\n", "max_recall_1 = max(test_metrics_df[\"Recall_1\"])\n", "max_f1_1 = max(test_metrics_df[\"F1-Score_1\"])\n", "\n", "print_colored(f\"\\n{'Model':<9} {'Accuracy ':<11} {'Precision_0 ':<10} {'Recall_0 ':<11} {'F1-Score_0 ':<11} \"\n", " f\"{'Precision_1 ':<10} {'Recall_1 ':<11} {'F1-Score_1':<13}\", \"blue\")\n", "\n", "for i in range(len(test_metrics_df[\"Model\"])):\n", " print_colored(f\"{test_metrics_df['Model'][i]:<10}\", \"red\", end=\"\")\n", "\n", " accuracy = test_metrics_df[\"Accuracy\"][i]\n", " precision_0 = test_metrics_df[\"Precision_0\"][i]\n", " recall_0 = test_metrics_df[\"Recall_0\"][i]\n", " f1_0 = test_metrics_df[\"F1-Score_0\"][i]\n", "\n", " precision_1 = test_metrics_df[\"Precision_1\"][i]\n", " recall_1 = test_metrics_df[\"Recall_1\"][i]\n", " f1_1 = test_metrics_df[\"F1-Score_1\"][i]\n", "\n", " print_colored(f\"{accuracy:<12.4f}\", bg_color=\"blue\" if accuracy == max_accuracy else \"\", color=\"black\", end=\"\")\n", " print_colored(f\"{precision_0:<12.4f} \", bg_color=\"blue\" if precision_0 == max_precision_0 else \"\", color=\"black\", end=\"\")\n", " print_colored(f\"{recall_0:<12.4f} \", bg_color=\"blue\" if recall_0 == max_recall_0 else \"\", color=\"black\", end=\"\")\n", " print_colored(f\"{f1_0:<12.4f} \", bg_color=\"blue\" if f1_0 == max_f1_0 else \"\", color=\"black\", end=\"\")\n", "\n", " print_colored(f\"{precision_1:<12.4f} \", bg_color=\"blue\" if precision_1 == max_precision_1 else \"\", color=\"black\", end=\"\")\n", " print_colored(f\"{recall_1:<12.4f} \", bg_color=\"blue\" if recall_1 == max_recall_1 else \"\", color=\"black\", end=\"\")\n", " print_colored(f\"{f1_1:<12.4f}\", bg_color=\"blue\" if f1_1 == max_f1_1 else \"\", color=\"black\", end=\"\\n\")\n" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "name": "python" } }, "nbformat": 4, "nbformat_minor": 5 }