telegram-analytics / stylometry.py
rottg's picture
Update code
a922175 verified
"""
Advanced Stylometry Analysis Module for Hebrew Text
Detects potential duplicate accounts based on writing style patterns.
Uses:
- sentence-transformers for Hebrew embeddings (writing style fingerprint)
- scikit-learn for DBSCAN clustering + TF-IDF on function words
- Hebrew-specific linguistic features (gender, formality, slang)
"""
import re
import sqlite3
import math
import pickle
import os
from collections import Counter, defaultdict
from datetime import datetime, timedelta
from typing import Dict, List, Tuple, Optional, Set
import numpy as np
# ==========================================
# HEBREW LINGUISTIC PATTERNS
# ==========================================
# Hebrew character ranges
HEBREW_PATTERN = re.compile(r'[\u0590-\u05FF]')
ENGLISH_PATTERN = re.compile(r'[a-zA-Z]')
EMOJI_PATTERN = re.compile(
"["
"\U0001F600-\U0001F64F" # emoticons
"\U0001F300-\U0001F5FF" # symbols & pictographs
"\U0001F680-\U0001F6FF" # transport & map symbols
"\U0001F1E0-\U0001F1FF" # flags
"\U00002702-\U000027B0"
"\U000024C2-\U0001F251"
"]+",
flags=re.UNICODE
)
# Hebrew function words (high frequency, style indicators)
HEBREW_FUNCTION_WORDS = [
'砖诇', '讗转', '注诇', '注诐', '讗诇', '诪谉', '讘讬谉', '诇驻谞讬', '讗讞专讬', '转讞转',
'讗谞讬', '讗转讛', '讗转', '讛讜讗', '讛讬讗', '讗谞讞谞讜', '讗转诐', '讗转谉', '讛诐', '讛谉',
'讝讛', '讝讗转', '讝讜', '讗诇讛', '讗诇讜',
'讻讬', '讗诐', '讗讜', '讙诐', '专拽', '讗讘诇', '讗诇讗', '诇诪专讜转', '讘讙诇诇', '讻讚讬',
'诪讛', '诪讬', '讗讬驻讛', '诪转讬', '诇诪讛', '讗讬讱', '讻诪讛',
'讻诇', '讛专讘讛', '拽爪转', '诪讗讜讚', '讬讜转专', '驻讞讜转', '讻诪讜',
'诇讗', '讻谉', '讗讬谉', '讬砖', '讛讬讛', '诇讛讬讜转', '注讜讚', '讻讘专',
]
# Formal vs informal markers
FORMAL_MARKERS = ['讗谞讜讻讬', '讛谞谞讬', '注诇讬讻诐', '讘讘拽砖讛', '转讜讚讛 专讘讛', '讘讻讘讜讚 专讘', '诇讻讘讜讚']
INFORMAL_MARKERS = ['讗讞讬', '讙讘专', '讗讞诇讛', '住讘讘讛', '讬讗诇诇讛', '讜讜讗诇讛', '讘讗住讛', '讞讞讞讞', '讞讞讞', '诇讜诇', 'wtf', 'omg']
# Hebrew slang and expressions
HEBREW_SLANG = [
'讗讞诇讛', '住讘讘讛', '讬讗诇诇讛', '讜讜讗诇讛', '讘讗住讛', '讞讘诇', '诪讙谞讬讘', '讗砖讻专讛',
'讞讞讞讞', '讞讞讞', '讛讛讛讛', '诪诪诪诪', '讗讛讛讛', '谞讜', '讟讜讘', '讘住讚专',
'驻讬爪讜抓', '诪砖讛讜', '讻讗讬诇讜', '住转诐', '诪诪砖', '驻砖讜讟', '谞讜专讗', '诪诇讗',
]
# Hebrew acronyms
HEBREW_ACRONYMS = ['讘注讝讛砖', '讗讻讗', '谞诇注谞讚', '转谞爪讘讛', '讝爪诇', '讘住"讚', '讘注"讛', '讗讬"讛', '讘诇"谞']
# Gender markers in verbs (past tense patterns)
MALE_VERB_ENDINGS = ['转讬', '转', '谞讜', '转诐'] # 讛诇讻转讬, 讛诇讻转, 讛诇讻谞讜
FEMALE_VERB_ENDINGS = ['转讬', '转', '谞讜', '转谉'] # 讛诇讻转讬, 讛诇讻转 (female), 讛诇讻谞讜
# Repeated character pattern (emotional expression)
REPEATED_CHARS_PATTERN = re.compile(r'(.)\1{2,}')
# Word with numbers pattern (l33t speak)
LEET_PATTERN = re.compile(r'\b\w*\d+\w*\b')
class AdvancedStyleFeatures:
"""Enhanced features extracted from a user's messages."""
def __init__(self, user_id: str, user_name: str):
self.user_id = user_id
self.user_name = user_name
self.message_count = 0
# === Basic Statistics ===
self.avg_message_length = 0.0
self.std_message_length = 0.0
self.avg_word_length = 0.0
self.avg_words_per_message = 0.0
# === Character Ratios ===
self.hebrew_ratio = 0.0
self.english_ratio = 0.0
self.digit_ratio = 0.0
self.emoji_ratio = 0.0
self.punctuation_ratio = 0.0
# === Punctuation Patterns ===
self.comma_rate = 0.0
self.period_rate = 0.0
self.question_rate = 0.0
self.exclamation_rate = 0.0
self.ellipsis_rate = 0.0
self.quote_rate = 0.0
# === Hebrew-Specific Features ===
self.formality_score = 0.0 # -1 (informal) to +1 (formal)
self.slang_rate = 0.0
self.acronym_rate = 0.0
self.repeated_chars_rate = 0.0
self.leet_speak_rate = 0.0
# === Linguistic Patterns ===
self.function_word_freq: Dict[str, float] = {}
self.unique_word_ratio = 0.0
self.hapax_ratio = 0.0 # Words used only once
self.short_message_ratio = 0.0
self.long_message_ratio = 0.0
# === Time Patterns ===
self.hour_distribution = np.zeros(24)
self.weekday_distribution = np.zeros(7)
self.weekend_ratio = 0.0
self.night_owl_ratio = 0.0 # Messages between 00:00-06:00
# === Response Patterns ===
self.reply_rate = 0.0
self.avg_response_words = 0.0
# === N-gram Features ===
self.char_bigrams: Dict[str, float] = {}
self.char_trigrams: Dict[str, float] = {}
self.word_bigrams: Dict[str, float] = {}
# === Embedding (from sentence-transformers) ===
self.style_embedding: Optional[np.ndarray] = None
# === TF-IDF Vector ===
self.tfidf_vector: Optional[np.ndarray] = None
# === Combined Feature Vector ===
self.feature_vector: Optional[np.ndarray] = None
def to_dict(self) -> dict:
return {
'user_id': self.user_id,
'user_name': self.user_name,
'message_count': self.message_count,
'avg_message_length': round(self.avg_message_length, 2),
'avg_word_length': round(self.avg_word_length, 2),
'hebrew_ratio': round(self.hebrew_ratio, 3),
'english_ratio': round(self.english_ratio, 3),
'emoji_ratio': round(self.emoji_ratio, 4),
'formality_score': round(self.formality_score, 2),
'slang_rate': round(self.slang_rate, 3),
'question_rate': round(self.question_rate, 3),
'exclamation_rate': round(self.exclamation_rate, 3),
'repeated_chars_rate': round(self.repeated_chars_rate, 3),
'weekend_ratio': round(self.weekend_ratio, 3),
'night_owl_ratio': round(self.night_owl_ratio, 3),
'unique_word_ratio': round(self.unique_word_ratio, 3),
}
class AdvancedStylometryAnalyzer:
"""
ML-powered stylometry analyzer using:
- sentence-transformers for Hebrew writing style embeddings
- scikit-learn for TF-IDF and DBSCAN clustering
- Hebrew linguistic feature extraction
"""
def __init__(self, db_path: str = 'telegram.db'):
self.db_path = db_path
self.user_features: Dict[int, AdvancedStyleFeatures] = {}
self.similarity_threshold = 0.85
# ML components (lazy loaded)
self._embedding_model = None
self._tfidf_vectorizer = None
self._scaler = None
# Cache directory
self.cache_dir = os.path.dirname(os.path.abspath(__file__))
@property
def embedding_model(self):
"""Lazy load sentence-transformers model."""
if self._embedding_model is None:
try:
from sentence_transformers import SentenceTransformer
# Use multilingual model that supports Hebrew well
# Alternative: 'imvladikon/sentence-transformers-alephbert' for pure Hebrew
print("Loading Hebrew embedding model...")
self._embedding_model = SentenceTransformer('paraphrase-multilingual-MiniLM-L12-v2')
print("Embedding model loaded.")
except Exception as e:
print(f"Could not load embedding model: {e}")
self._embedding_model = False # Mark as failed
return self._embedding_model if self._embedding_model else None
def get_active_users(self, min_messages: int = 300, days: int = 365) -> List[Tuple[str, str, int]]:
"""Get users active in the last N days with at least min_messages."""
cutoff_date = datetime.now() - timedelta(days=days)
cutoff_timestamp = int(cutoff_date.timestamp())
conn = sqlite3.connect(self.db_path)
cursor = conn.cursor()
# Use from_id and from_name directly from messages table
query = """
SELECT from_id, MAX(from_name) as name, COUNT(*) as msg_count
FROM messages
WHERE date_unixtime >= ?
AND from_id IS NOT NULL
AND text_plain IS NOT NULL
AND text_plain != ''
GROUP BY from_id
HAVING msg_count >= ?
ORDER BY msg_count DESC
"""
cursor.execute(query, (cutoff_timestamp, min_messages))
users = cursor.fetchall()
conn.close()
return users
def get_user_messages(self, user_id: str, days: int = 365) -> List[Tuple[str, str]]:
"""Get messages for a user (text, date)."""
cutoff_date = datetime.now() - timedelta(days=days)
cutoff_timestamp = int(cutoff_date.timestamp())
conn = sqlite3.connect(self.db_path)
cursor = conn.cursor()
query = """
SELECT text_plain, date FROM messages
WHERE from_id = ? AND date_unixtime >= ?
AND text_plain IS NOT NULL AND text_plain != ''
ORDER BY date_unixtime
"""
cursor.execute(query, (user_id, cutoff_timestamp))
messages = cursor.fetchall()
conn.close()
return messages
def extract_features(self, user_id: str, user_name: str,
messages: List[Tuple[str, str]]) -> AdvancedStyleFeatures:
"""Extract comprehensive stylometric features from user messages."""
features = AdvancedStyleFeatures(user_id, user_name)
features.message_count = len(messages)
if not messages:
return features
# Collect all text for analysis
all_texts = [msg[0] for msg in messages if msg[0]]
all_text_combined = ' '.join(all_texts)
# === Basic Statistics ===
message_lengths = [len(text) for text in all_texts]
features.avg_message_length = np.mean(message_lengths)
features.std_message_length = np.std(message_lengths)
all_words = []
word_counts_per_msg = []
for text in all_texts:
words = text.split()
all_words.extend(words)
word_counts_per_msg.append(len(words))
if all_words:
word_lengths = [len(w) for w in all_words]
features.avg_word_length = np.mean(word_lengths)
features.avg_words_per_message = np.mean(word_counts_per_msg)
# === Character Ratios ===
total_chars = len(all_text_combined)
if total_chars > 0:
hebrew_chars = len(HEBREW_PATTERN.findall(all_text_combined))
english_chars = len(ENGLISH_PATTERN.findall(all_text_combined))
digit_chars = sum(1 for c in all_text_combined if c.isdigit())
punct_chars = sum(1 for c in all_text_combined if c in '.,!?;:()[]{}')
emoji_count = len(EMOJI_PATTERN.findall(all_text_combined))
features.hebrew_ratio = hebrew_chars / total_chars
features.english_ratio = english_chars / total_chars
features.digit_ratio = digit_chars / total_chars
features.punctuation_ratio = punct_chars / total_chars
features.emoji_ratio = emoji_count / total_chars
# === Punctuation Patterns ===
n_msgs = len(messages)
features.comma_rate = all_text_combined.count(',') / n_msgs
features.period_rate = all_text_combined.count('.') / n_msgs
features.question_rate = all_text_combined.count('?') / n_msgs
features.exclamation_rate = all_text_combined.count('!') / n_msgs
features.ellipsis_rate = all_text_combined.count('...') / n_msgs
features.quote_rate = (all_text_combined.count('"') + all_text_combined.count("'")) / n_msgs
# === Hebrew-Specific Features ===
text_lower = all_text_combined.lower()
# Formality score
formal_count = sum(1 for marker in FORMAL_MARKERS if marker in all_text_combined)
informal_count = sum(1 for marker in INFORMAL_MARKERS if marker in text_lower)
total_markers = formal_count + informal_count
if total_markers > 0:
features.formality_score = (formal_count - informal_count) / total_markers
# Slang rate
slang_count = sum(1 for text in all_texts for slang in HEBREW_SLANG if slang in text)
features.slang_rate = slang_count / n_msgs
# Acronym rate
acronym_count = sum(1 for text in all_texts for acr in HEBREW_ACRONYMS if acr in text)
features.acronym_rate = acronym_count / n_msgs
# Repeated characters (emotional expression like 讞讞讞讞)
repeated_msgs = sum(1 for text in all_texts if REPEATED_CHARS_PATTERN.search(text))
features.repeated_chars_rate = repeated_msgs / n_msgs
# Leet speak rate
leet_count = sum(len(LEET_PATTERN.findall(text)) for text in all_texts)
features.leet_speak_rate = leet_count / n_msgs
# === Linguistic Patterns ===
# Function word frequency
word_counter = Counter(w.lower() for w in all_words)
total_words = len(all_words)
for fw in HEBREW_FUNCTION_WORDS:
features.function_word_freq[fw] = word_counter.get(fw, 0) / max(1, total_words)
# Vocabulary richness
unique_words = set(w.lower() for w in all_words)
features.unique_word_ratio = len(unique_words) / max(1, total_words)
# Hapax legomena (words appearing only once)
hapax_count = sum(1 for w, c in word_counter.items() if c == 1)
features.hapax_ratio = hapax_count / max(1, len(unique_words))
# Message length categories
features.short_message_ratio = sum(1 for wc in word_counts_per_msg if wc < 5) / n_msgs
features.long_message_ratio = sum(1 for wc in word_counts_per_msg if wc > 30) / n_msgs
# === Time Patterns ===
hour_counts = np.zeros(24)
weekday_counts = np.zeros(7)
night_msgs = 0
weekend_msgs = 0
for text, date_str in messages:
try:
if 'T' in date_str:
dt = datetime.fromisoformat(date_str.replace('Z', '+00:00'))
else:
dt = datetime.strptime(date_str[:19], '%Y-%m-%d %H:%M:%S')
hour_counts[dt.hour] += 1
weekday_counts[dt.weekday()] += 1
if 0 <= dt.hour < 6:
night_msgs += 1
if dt.weekday() >= 5: # Saturday=5, Sunday=6
weekend_msgs += 1
except:
pass
# Normalize
if hour_counts.sum() > 0:
features.hour_distribution = hour_counts / hour_counts.sum()
if weekday_counts.sum() > 0:
features.weekday_distribution = weekday_counts / weekday_counts.sum()
features.weekend_ratio = weekend_msgs / n_msgs
features.night_owl_ratio = night_msgs / n_msgs
# === N-gram Features ===
# Character bigrams
char_bigram_counter = Counter()
for text in all_texts:
clean_text = re.sub(r'\s+', ' ', text.lower())
for i in range(len(clean_text) - 1):
bg = clean_text[i:i+2]
if bg.strip():
char_bigram_counter[bg] += 1
total_bigrams = sum(char_bigram_counter.values())
if total_bigrams > 0:
for bg, count in char_bigram_counter.most_common(100):
features.char_bigrams[bg] = count / total_bigrams
# Character trigrams
char_trigram_counter = Counter()
for text in all_texts:
clean_text = re.sub(r'\s+', ' ', text.lower())
for i in range(len(clean_text) - 2):
tg = clean_text[i:i+3]
if tg.strip():
char_trigram_counter[tg] += 1
total_trigrams = sum(char_trigram_counter.values())
if total_trigrams > 0:
for tg, count in char_trigram_counter.most_common(100):
features.char_trigrams[tg] = count / total_trigrams
# Word bigrams
word_bigram_counter = Counter()
for text in all_texts:
words = text.lower().split()
for i in range(len(words) - 1):
wb = f"{words[i]} {words[i+1]}"
word_bigram_counter[wb] += 1
total_word_bigrams = sum(word_bigram_counter.values())
if total_word_bigrams > 0:
for wb, count in word_bigram_counter.most_common(50):
features.word_bigrams[wb] = count / total_word_bigrams
# === Generate Style Embedding ===
if self.embedding_model:
try:
# Sample messages for embedding (limit for performance)
sample_texts = all_texts[:100] if len(all_texts) > 100 else all_texts
# Combine into a style sample
style_sample = ' '.join(sample_texts)[:5000] # Limit length
features.style_embedding = self.embedding_model.encode(style_sample, show_progress_bar=False)
except Exception as e:
print(f"Embedding error for user {user_id}: {e}")
# === Build Numeric Feature Vector ===
features.feature_vector = self._build_feature_vector(features)
return features
def _build_feature_vector(self, f: AdvancedStyleFeatures) -> np.ndarray:
"""Build normalized feature vector for similarity comparison."""
vector = [
# Basic stats (normalized)
f.avg_message_length / 200,
f.std_message_length / 100,
f.avg_word_length / 10,
f.avg_words_per_message / 20,
# Character ratios
f.hebrew_ratio,
f.english_ratio,
f.digit_ratio * 10,
f.emoji_ratio * 100,
f.punctuation_ratio * 10,
# Punctuation patterns
f.comma_rate / 2,
f.period_rate / 2,
f.question_rate,
f.exclamation_rate,
f.ellipsis_rate * 5,
f.quote_rate,
# Hebrew-specific
f.formality_score,
f.slang_rate * 5,
f.acronym_rate * 10,
f.repeated_chars_rate * 5,
f.leet_speak_rate * 10,
# Linguistic
f.unique_word_ratio,
f.hapax_ratio,
f.short_message_ratio,
f.long_message_ratio,
# Time patterns
f.weekend_ratio,
f.night_owl_ratio * 5,
]
# Add hour distribution (24 values)
vector.extend(f.hour_distribution.tolist())
# Add weekday distribution (7 values)
vector.extend(f.weekday_distribution.tolist())
# Add top function word frequencies (20 values)
for fw in HEBREW_FUNCTION_WORDS[:20]:
vector.append(f.function_word_freq.get(fw, 0) * 100)
return np.array(vector)
def calculate_similarity(self, f1: AdvancedStyleFeatures, f2: AdvancedStyleFeatures) -> Tuple[float, Dict]:
"""
Calculate comprehensive similarity between two users.
Returns overall score and component breakdown.
"""
scores = {}
# 1. Feature vector similarity (cosine)
if f1.feature_vector is not None and f2.feature_vector is not None:
v1, v2 = f1.feature_vector, f2.feature_vector
dot_product = np.dot(v1, v2)
norm1, norm2 = np.linalg.norm(v1), np.linalg.norm(v2)
if norm1 > 0 and norm2 > 0:
scores['feature_cosine'] = float(dot_product / (norm1 * norm2))
else:
scores['feature_cosine'] = 0.0
else:
scores['feature_cosine'] = 0.0
# 2. Embedding similarity (if available)
if f1.style_embedding is not None and f2.style_embedding is not None:
e1, e2 = f1.style_embedding, f2.style_embedding
dot_product = np.dot(e1, e2)
norm1, norm2 = np.linalg.norm(e1), np.linalg.norm(e2)
if norm1 > 0 and norm2 > 0:
scores['embedding_cosine'] = float(dot_product / (norm1 * norm2))
else:
scores['embedding_cosine'] = 0.0
else:
scores['embedding_cosine'] = None
# 3. Character bigram similarity (Jaccard-like)
scores['bigram_overlap'] = self._ngram_similarity(f1.char_bigrams, f2.char_bigrams)
# 4. Trigram similarity
scores['trigram_overlap'] = self._ngram_similarity(f1.char_trigrams, f2.char_trigrams)
# 5. Word bigram similarity
scores['word_bigram_overlap'] = self._ngram_similarity(f1.word_bigrams, f2.word_bigrams)
# 6. Time pattern similarity (hour distribution)
if f1.hour_distribution.sum() > 0 and f2.hour_distribution.sum() > 0:
scores['time_pattern'] = float(np.dot(f1.hour_distribution, f2.hour_distribution))
else:
scores['time_pattern'] = 0.0
# === Threshold-based scoring ===
# Feature Vector is the most reliable discriminator. Use it as a gate:
# - Below 94%: heavy penalty (likely different people)
# - 94-96%: moderate score
# - Above 96%: bonus (likely same person)
feature_score = scores['feature_cosine']
bigram_score = scores['bigram_overlap']
# Base score from key metrics (feature vector is primary)
base_score = (
feature_score * 0.50 +
bigram_score * 0.30 +
scores['trigram_overlap'] * 0.10 +
(scores['embedding_cosine'] * 0.10 if scores['embedding_cosine'] is not None else 0)
)
# Apply threshold-based multipliers
if feature_score >= 0.96:
# Very high feature similarity - likely same person
multiplier = 1.15
elif feature_score >= 0.94:
# High similarity - possible match
multiplier = 1.0
elif feature_score >= 0.90:
# Moderate similarity - penalize
multiplier = 0.75
else:
# Low similarity - heavy penalty
multiplier = 0.5
# Additional penalty if bigrams are low
if bigram_score < 0.80:
multiplier *= 0.85
elif bigram_score >= 0.85:
multiplier *= 1.05
overall = base_score * multiplier
# Cap at 100%
overall = min(overall, 1.0)
return overall, scores
def _ngram_similarity(self, ng1: Dict[str, float], ng2: Dict[str, float]) -> float:
"""Calculate similarity between n-gram distributions."""
if not ng1 or not ng2:
return 0.0
all_ngrams = set(ng1.keys()) | set(ng2.keys())
if not all_ngrams:
return 0.0
intersection = 0.0
union = 0.0
for ng in all_ngrams:
v1 = ng1.get(ng, 0)
v2 = ng2.get(ng, 0)
intersection += min(v1, v2)
union += max(v1, v2)
if union == 0:
return 0.0
return intersection / union
def cluster_users(self, min_cluster_size: int = 2) -> List[List[int]]:
"""
Use DBSCAN to automatically cluster users with similar writing styles.
Returns list of clusters (each cluster is a list of user_ids).
"""
if len(self.user_features) < 2:
return []
try:
from sklearn.cluster import DBSCAN
from sklearn.preprocessing import StandardScaler
except ImportError:
print("scikit-learn not available for clustering")
return []
# Build feature matrix
user_ids = list(self.user_features.keys())
feature_matrix = []
for uid in user_ids:
f = self.user_features[uid]
if f.feature_vector is not None:
# Combine feature vector with embedding if available
if f.style_embedding is not None:
combined = np.concatenate([f.feature_vector, f.style_embedding])
else:
combined = f.feature_vector
feature_matrix.append(combined)
else:
feature_matrix.append(np.zeros(50)) # Fallback
feature_matrix = np.array(feature_matrix)
# Normalize features
scaler = StandardScaler()
features_scaled = scaler.fit_transform(feature_matrix)
# DBSCAN clustering
# eps: maximum distance between samples in a cluster
# min_samples: minimum samples to form a cluster
dbscan = DBSCAN(eps=0.5, min_samples=min_cluster_size, metric='cosine')
labels = dbscan.fit_predict(features_scaled)
# Group users by cluster
clusters = defaultdict(list)
for i, label in enumerate(labels):
if label >= 0: # -1 means noise (no cluster)
clusters[label].append(user_ids[i])
return [users for users in clusters.values() if len(users) >= min_cluster_size]
def analyze_all_users(self, min_messages: int = 300, days: int = 365,
progress_callback=None) -> Dict:
"""Analyze all active users and find potential duplicates."""
# Get active users
users = self.get_active_users(min_messages, days)
total_users = len(users)
if progress_callback:
progress_callback('users_found', total_users)
# Extract features for each user
self.user_features = {}
for idx, (user_id, user_name, msg_count) in enumerate(users):
messages = self.get_user_messages(user_id, days)
features = self.extract_features(user_id, user_name or f"User_{user_id}", messages)
self.user_features[user_id] = features
if progress_callback:
progress_callback('user_processed', idx + 1, total_users, user_name or f"User_{user_id}")
# Find similar pairs
if progress_callback:
progress_callback('comparing', 0)
similar_pairs = []
user_ids = list(self.user_features.keys())
total_comparisons = len(user_ids) * (len(user_ids) - 1) // 2
comparison_count = 0
for i in range(len(user_ids)):
for j in range(i + 1, len(user_ids)):
uid1, uid2 = user_ids[i], user_ids[j]
f1, f2 = self.user_features[uid1], self.user_features[uid2]
similarity, score_breakdown = self.calculate_similarity(f1, f2)
if similarity >= self.similarity_threshold:
similar_pairs.append({
'user1': f1.to_dict(),
'user2': f2.to_dict(),
'similarity': round(similarity * 100, 1),
'scores': {k: round(v * 100, 1) if v is not None else None
for k, v in score_breakdown.items()},
'details': self._get_similarity_details(f1, f2, score_breakdown)
})
comparison_count += 1
if progress_callback and comparison_count % 100 == 0:
progress_callback('comparing', comparison_count, total_comparisons)
# Sort by similarity (highest first)
similar_pairs.sort(key=lambda x: x['similarity'], reverse=True)
# Run clustering
clusters = self.cluster_users(min_cluster_size=2)
cluster_info = []
for cluster in clusters:
cluster_users = [self.user_features[uid].to_dict() for uid in cluster]
cluster_info.append({
'users': cluster_users,
'size': len(cluster)
})
return {
'total_users_analyzed': total_users,
'threshold': self.similarity_threshold * 100,
'potential_duplicates': len(similar_pairs),
'pairs': similar_pairs,
'clusters': cluster_info,
'all_users': [f.to_dict() for f in self.user_features.values()],
'embedding_model_used': self.embedding_model is not None,
}
def _get_similarity_details(self, f1: AdvancedStyleFeatures, f2: AdvancedStyleFeatures,
scores: Dict) -> List[str]:
"""Get human-readable similarity details in Hebrew."""
details = []
# High embedding similarity
if scores.get('embedding_cosine') and scores['embedding_cosine'] > 0.85:
details.append("住讙谞讜谉 讻转讬讘讛 讚讜诪讛 诪讗讜讚 (AI embedding)")
# Message length
len_diff = abs(f1.avg_message_length - f2.avg_message_length)
if len_diff < 15:
details.append(f"讗讜专讱 讛讜讚注讛 讚讜诪讛 ({f1.avg_message_length:.0f} vs {f2.avg_message_length:.0f})")
# Hebrew/English ratio
heb_diff = abs(f1.hebrew_ratio - f2.hebrew_ratio)
if heb_diff < 0.1:
details.append(f"讬讞住 注讘专讬转 讚讜诪讛 ({f1.hebrew_ratio:.0%} vs {f2.hebrew_ratio:.0%})")
# Emoji usage
emoji_diff = abs(f1.emoji_ratio - f2.emoji_ratio)
if emoji_diff < 0.005 and (f1.emoji_ratio > 0.001 or f2.emoji_ratio > 0.001):
details.append("砖讬诪讜砖 讚讜诪讛 讘讗讬诪讜讙'讬")
# Formality
form_diff = abs(f1.formality_score - f2.formality_score)
if form_diff < 0.3:
if f1.formality_score > 0.3:
details.append("砖谞讬讛诐 讻讜转讘讬诐 讘住讙谞讜谉 驻讜专诪诇讬")
elif f1.formality_score < -0.3:
details.append("砖谞讬讛诐 讻讜转讘讬诐 讘住讙谞讜谉 诇讗 驻讜专诪诇讬")
# Slang usage
if abs(f1.slang_rate - f2.slang_rate) < 0.1:
if f1.slang_rate > 0.2:
details.append("砖讬诪讜砖 讚讜诪讛 讘住诇谞讙")
# Repeated characters
if abs(f1.repeated_chars_rate - f2.repeated_chars_rate) < 0.05:
if f1.repeated_chars_rate > 0.1:
details.append("砖谞讬讛诐 诪砖转诪砖讬诐 讘转讜讜讬诐 讞讜讝专讬诐 (讻诪讜 讞讞讞讞)")
# Time patterns
if scores.get('time_pattern', 0) > 0.8:
details.append("讚驻讜住 砖注讜转 驻注讬诇讜转 讚讜诪讛 诪讗讜讚")
# Weekend activity
weekend_diff = abs(f1.weekend_ratio - f2.weekend_ratio)
if weekend_diff < 0.1:
details.append("驻注讬诇讜转 讚讜诪讛 讘住讜驻\"砖")
# Night owl
if abs(f1.night_owl_ratio - f2.night_owl_ratio) < 0.05:
if f1.night_owl_ratio > 0.1:
details.append("砖谞讬讛诐 驻注讬诇讬诐 讘砖注讜转 讛诇讬诇讛")
# N-gram overlap
if scores.get('bigram_overlap', 0) > 0.6:
details.append("讚驻讜住讬 讗讜转讬讜转 讚讜诪讬诐 诪讗讜讚")
if scores.get('word_bigram_overlap', 0) > 0.4:
details.append("爪讬专讜驻讬 诪讬诇讬诐 讚讜诪讬诐")
return details
# Singleton instance
_analyzer_instance: Optional[AdvancedStylometryAnalyzer] = None
def get_stylometry_analyzer() -> AdvancedStylometryAnalyzer:
"""Get or create the stylometry analyzer singleton."""
global _analyzer_instance
if _analyzer_instance is None:
_analyzer_instance = AdvancedStylometryAnalyzer()
return _analyzer_instance