File size: 22,861 Bytes
c293f7c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
#!/usr/bin/env python3
"""
Advanced ML Classifier for Misinformation Detection
- Loads training data from structured CSV dataset (760+ examples)
- Multiple algorithms (Naive Bayes, SVM, Random Forest, Logistic Regression, Gradient Boosting)
- Feature engineering with TF-IDF, linguistic features, and Indian-context features
- Hinglish-aware with expanded fake news patterns
- Cross-validation, per-class metrics, and performance reporting
"""

import pandas as pd
import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.naive_bayes import MultinomialNB
from sklearn.svm import SVC
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import (
    RandomForestClassifier, GradientBoostingClassifier, VotingClassifier
)
from sklearn.pipeline import Pipeline, FeatureUnion
from sklearn.model_selection import train_test_split, cross_val_score, StratifiedKFold
from sklearn.metrics import (
    classification_report, confusion_matrix, accuracy_score,
    f1_score, roc_auc_score
)
from sklearn.base import BaseEstimator, TransformerMixin
from sklearn.preprocessing import MinMaxScaler
import re
import os
import json
import nltk
from nltk.sentiment import SentimentIntensityAnalyzer
import pickle
import logging
from pathlib import Path
from datetime import datetime

# Setup logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Ensure NLTK data is available
try:
    nltk.data.find('sentiment/vader_lexicon.zip')
except LookupError:
    nltk.download('vader_lexicon', quiet=True)


class LinguisticFeatureExtractor(BaseEstimator, TransformerMixin):
    """Extract linguistic features that indicate misinformation"""

    def __init__(self):
        self.sia = SentimentIntensityAnalyzer()

    def fit(self, X, y=None):
        return self

    def transform(self, X):
        features = []

        for text in X:
            text_lower = text.lower()
            words = text.split()
            word_count = max(len(words), 1)

            # Feature 1: Sensational keywords count (normalized)
            sensational_keywords = [
                'breaking', 'urgent', 'shocking', 'exclusive', 'exposed', 'revealed',
                'secret', 'hidden', 'conspiracy', 'cover-up', 'scandal', 'bombshell',
                'viral', 'trending', 'must see', 'unbelievable', 'incredible', 'amazing',
                'alert', 'proof', 'leaked', 'caught', 'busted', 'warning',
                # Additional sensational patterns
                'suppressed', 'banned', 'deleted', 'censored', 'silenced',
                'whistleblower', 'classified', 'blackout', 'hidden camera',
                'insider reveals', 'mainstream media hiding', 'doctors hate',
                'forward before deleted', 'share before removed'
            ]
            sensational_count = sum(1 for kw in sensational_keywords if kw in text_lower)
            sensational_ratio = sensational_count / word_count

            # Feature 2: Emotional manipulation words (normalized)
            emotional_words = [
                'outraged', 'furious', 'devastated', 'terrified', 'shocked',
                'disgusted', 'betrayed', 'abandoned', 'threatened', 'angry',
                'horrifying', 'nightmare', 'deadly', 'catastrophic', 'crisis'
            ]
            emotional_count = sum(1 for w in emotional_words if w in text_lower)
            emotional_ratio = emotional_count / word_count

            # Feature 3: Attribution indicators (binary)
            attribution_indicators = [
                'according to', 'sources say', 'officials confirm', 'study shows',
                'research indicates', 'data reveals', 'experts believe', 'report states',
                'published in', 'peer-reviewed', 'government says', 'ministry of',
                'official statement', 'press release', 'data from', 'survey finds'
            ]
            has_attribution = 1.0 if any(ind in text_lower for ind in attribution_indicators) else 0.0

            # Feature 4: Excessive punctuation ratios
            exclamation_ratio = text.count('!') / word_count
            question_ratio = text.count('?') / word_count
            ellipsis_count = text.count('...') / word_count

            # Feature 5: ALL CAPS ratio
            caps_words = sum(1 for w in words if w.isupper() and len(w) > 2)
            caps_ratio = caps_words / word_count

            # Feature 6: Sentiment extremity (VADER)
            sentiment = self.sia.polarity_scores(text)
            sentiment_extremity = abs(sentiment['compound'])
            negative_sentiment = abs(sentiment['neg'])

            # Feature 7: Text length features
            text_length = len(text)
            avg_word_length = sum(len(w) for w in words) / word_count
            # Very short texts are more suspicious (WhatsApp forwards)
            is_very_short = 1.0 if text_length < 100 else 0.0

            # Feature 8: Clickbait patterns
            clickbait_patterns = [
                r'\d+ (things|ways|reasons|facts)',
                r'you (won\'t|will not) believe',
                r'this (will|might) (shock|surprise) you',
                r'number \d+ will',
                r'what happens next',
                r'doctors hate',
                r'one (simple|weird) trick'
            ]
            clickbait_count = sum(1 for p in clickbait_patterns if re.search(p, text_lower))

            # Feature 9: Urgency language
            urgency_words = [
                'now', 'immediately', 'right now', 'hurry', 'before it\'s too late',
                'act fast', 'don\'t wait', 'last chance', 'share before deleted',
                'forward to', 'spread the word', 'wake up'
            ]
            urgency_count = sum(1 for u in urgency_words if u in text_lower)

            # Feature 10: Numerical claim density
            number_pattern = re.findall(r'\d+', text)
            number_density = len(number_pattern) / word_count

            # Feature 11: Source credibility signals (lack of)
            no_source = 1.0 if not any(sig in text_lower for sig in [
                'says', 'said', 'reported', 'confirmed', 'announced',
                'stated', 'according', 'ministry', 'department', 'official'
            ]) else 0.0

            features.append([
                sensational_count,
                sensational_ratio,
                emotional_count,
                emotional_ratio,
                has_attribution,
                exclamation_ratio,
                question_ratio,
                ellipsis_count,
                caps_ratio,
                sentiment_extremity,
                negative_sentiment,
                is_very_short,
                avg_word_length,
                clickbait_count,
                urgency_count,
                number_density,
                no_source
            ])

        return np.array(features)


class IndianContextFeatureExtractor(BaseEstimator, TransformerMixin):
    """Extract India-specific features that indicate misinformation"""

    def fit(self, X, y=None):
        return self

    def transform(self, X):
        features = []

        for text in X:
            text_lower = text.lower()
            words = text.split()
            word_count = max(len(words), 1)

            # Feature 1: Hindi/Hinglish sensational keywords
            hindi_sensational = [
                'khatre mein', 'sabko batao', 'sach samne aaya', 'janta jaago',
                'share karo', 'forward karo', 'exposed', 'parda faash',
                'sarkaar', 'chupa raha', 'media blackout', 'sachchi news',
                'ye dekho', 'video dekho', 'photo dekho', 'sabko batao'
            ]
            hindi_sensational_count = sum(1 for kw in hindi_sensational if kw in text_lower)

            # Feature 2: WhatsApp forward patterns
            whatsapp_patterns = [
                'forward to', 'share to', 'send to', 'forward karo',
                '10 logo', '10 people', 'before deleted',
                'share before', 'ye message', 'group mein', 'sabko bhejo',
                'please share', 'must share', 'jarur share', 'zaroor forward'
            ]
            whatsapp_count = sum(1 for p in whatsapp_patterns if p in text_lower)

            # Feature 3: Communal trigger patterns
            communal_triggers = [
                'hindu khatre', 'muslim attack', 'love jihad', 'conversion',
                'temple destroy', 'mandir tod', 'mosque', 'masjid',
                'communal', 'riot', 'lynching', 'mob', 'religious violence',
                'anti-national', 'deshdrohi', 'traitor', 'jihad'
            ]
            communal_count = sum(1 for t in communal_triggers if t in text_lower)

            # Feature 4: Fake official/authority patterns
            fake_authority = [
                'government ne kaha', 'sarkar ka aadesh', 'pm ne bola',
                'who ne maana', 'doctor ne confirm', 'scientist ne bola',
                'nasa confirmed', 'un ne kaha', 'supreme court ne',
                'rbi ne bola', 'army ne kaha'
            ]
            fake_authority_count = sum(1 for f in fake_authority if f in text_lower)

            # Feature 5: Miracle/cure claims (India-specific)
            miracle_claims = [
                'cure', 'theek', 'ilaaj', 'upchar', 'ramban', 'miracle',
                'ayurvedic cure', 'desi nuskha', 'gharelu upay',
                'haldi', 'tulsi', 'neem', 'cow urine', 'gaumutra',
                'patanjali', 'baba ramdev',
                # Additional miracle/pseudoscience patterns
                'miracle cure', 'ancient remedy', 'natural cure', '100% effective',
                'doctors hiding', 'pharma hiding', 'permanently cures', 'guaranteed cure',
                'theek ho jata', 'cancer theek', 'dawai nahi chahiye', 'bilkul theek'
            ]
            miracle_count = sum(1 for m in miracle_claims if m in text_lower)

            # Feature 6: Inflated statistics (crore, lakh with large numbers)
            inflated_stats = len(re.findall(
                r'\d+\s*(crore|lakh|billion|million|percent|%)', text_lower
            ))

            # Feature 7: India-specific credibility signals
            indian_credible = [
                'pib', 'pti', 'ani', 'ians', 'press trust',
                'doordarshan', 'all india radio', 'prasar bharati',
                'ministry of', 'niti aayog', 'rbi', 'sebi',
                'isro', 'drdo', 'icar', 'icmr', 'aiims',
                'supreme court', 'high court', 'election commission'
            ]
            has_indian_source = 1.0 if any(s in text_lower for s in indian_credible) else 0.0

            # Feature 8: Deepfake/media manipulation indicators
            media_manipulation = [
                'old photo', 'old video', 'cropped', 'morphed', 'doctored',
                'fake photo', 'fake video', 'edited', 'photoshopped',
                'out of context', 'misleading', 'actually from', 'different country',
                'ai generated', 'deepfake'
            ]
            media_manip_count = sum(1 for m in media_manipulation if m in text_lower)

            features.append([
                hindi_sensational_count,
                whatsapp_count,
                communal_count,
                fake_authority_count,
                miracle_count,
                inflated_stats,
                has_indian_source,
                media_manip_count
            ])

        return np.array(features)


def load_training_data():
    """Load training data from CSV dataset file"""
    dataset_path = Path(__file__).parent / 'datasets' / 'indian_misinformation_v4.csv'

    if not dataset_path.exists():
        logger.warning(f"Dataset file not found at {dataset_path}, falling back to built-in data")
        return _create_fallback_training_data()

    try:
        df = pd.read_csv(dataset_path)

        # Validate required columns
        required_cols = ['text', 'label']
        if not all(col in df.columns for col in required_cols):
            raise ValueError(f"Dataset must contain columns: {required_cols}")

        # Clean data
        df = df.dropna(subset=['text', 'label'])
        df['text'] = df['text'].astype(str).str.strip()
        df['label'] = df['label'].astype(int)
        df = df[df['text'].str.len() > 10]  # Remove very short entries

        # Remove duplicates
        df = df.drop_duplicates(subset=['text'])

        fake_count = (df['label'] == 1).sum()
        real_count = (df['label'] == 0).sum()
        logger.info(f"Loaded dataset from {dataset_path}: {len(df)} examples "
                     f"(fake: {fake_count}, real: {real_count})")

        return df

    except Exception as e:
        logger.error(f"Failed to load dataset: {e}")
        return _create_fallback_training_data()


def _create_fallback_training_data():
    """Minimal fallback dataset if CSV is unavailable"""
    examples = [
        ("BREAKING: Government secretly controlling weather with 5G towers!", 1),
        ("EXPOSED: Hidden documents reveal vaccine microchips are real", 1),
        ("SHOCKING: Cow urine cures cancer - doctors hiding truth!", 1),
        ("ALERT: All bank accounts will be frozen tomorrow - withdraw now!", 1),
        ("Hindu khatre mein hai - 2025 tak India Islamic desh ban jayega", 1),
        ("VIRAL: Miracle herb cures diabetes in 3 days - pharma hates this", 1),
        ("URGENT: WhatsApp band hone wala hai - forward karo 10 logo ko", 1),
        ("EXPOSED: EVM machines rigged in elections - video proof!", 1),
        ("Share before deleted - secret plan of government exposed!", 1),
        ("BREAKING: Earthquake predicted for Delhi next week - evacuate!", 1),
        ("Government announces new infrastructure development plan for northeast India", 0),
        ("Parliament passes bill to improve healthcare access in rural areas", 0),
        ("According to RBI, GDP growth rate stands at 6.8 percent this quarter", 0),
        ("ISRO successfully launches communication satellite from Sriharikota", 0),
        ("Supreme Court delivers verdict on constitutional matter regarding privacy", 0),
        ("Medical study published in AIIMS journal shows effectiveness of new treatment", 0),
        ("Stock market shows steady growth this quarter on strong corporate earnings", 0),
        ("Weather department forecasts normal monsoon this year across India", 0),
        ("India signs bilateral fisheries agreement with Sri Lanka for sustainable fishing", 0),
        ("Fact-check: Viral claim about banning Rs 500 notes is FALSE says RBI", 0),
    ]
    df = pd.DataFrame(examples, columns=['text', 'label'])
    logger.warning(f"Using fallback dataset with only {len(df)} examples")
    return df


def build_advanced_classifier():
    """Build advanced ensemble classifier with multiple algorithms"""

    # Load training data from CSV
    df = load_training_data()

    # Split data
    X_train, X_test, y_train, y_test = train_test_split(
        df['text'], df['label'], test_size=0.2, random_state=42, stratify=df['label']
    )

    logger.info(f"Training set: {len(X_train)} examples")
    logger.info(f"Test set: {len(X_test)} examples")

    # --- Feature Engineering ---
    # TF-IDF on words: lowered min_df for better coverage
    tfidf_words = TfidfVectorizer(
        max_features=8000,
        stop_words='english',
        ngram_range=(1, 3),
        min_df=1,
        max_df=0.85,
        sublinear_tf=True
    )

    # TF-IDF on character n-grams: captures misspellings and transliteration
    tfidf_chars = TfidfVectorizer(
        analyzer='char_wb',
        ngram_range=(3, 6),
        max_features=3000,
        min_df=1,
        max_df=0.9,
        sublinear_tf=True
    )

    # Linguistic feature extractor
    linguistic_features = LinguisticFeatureExtractor()

    # Indian context feature extractor
    indian_features = IndianContextFeatureExtractor()

    # Combine all features
    feature_union = FeatureUnion([
        ('tfidf_words', tfidf_words),
        ('tfidf_chars', tfidf_chars),
        ('linguistic', linguistic_features),
        ('indian_context', indian_features)
    ])

    # --- Ensemble Classifier ---
    nb_classifier = MultinomialNB(alpha=0.05)
    svm_classifier = SVC(kernel='linear', probability=True, C=1.5, class_weight='balanced')
    rf_classifier = RandomForestClassifier(
        n_estimators=200, max_depth=20, random_state=42, class_weight='balanced'
    )
    lr_classifier = LogisticRegression(
        C=2.0, max_iter=1000, solver='lbfgs', class_weight='balanced'
    )
    gb_classifier = GradientBoostingClassifier(
        n_estimators=150, max_depth=5, learning_rate=0.1, random_state=42
    )

    # 5-model ensemble with soft voting
    ensemble_classifier = VotingClassifier([
        ('nb', nb_classifier),
        ('svm', svm_classifier),
        ('rf', rf_classifier),
        ('lr', lr_classifier),
        ('gb', gb_classifier)
    ], voting='soft', weights=[1, 2, 2, 2, 2])

    # Final pipeline
    classifier_pipeline = Pipeline([
        ('features', feature_union),
        ('classifier', ensemble_classifier)
    ])

    # --- Training ---
    logger.info("Training advanced 5-model ensemble classifier...")
    classifier_pipeline.fit(X_train, y_train)

    # --- Evaluation ---
    logger.info("Evaluating classifier performance...")

    # Cross-validation with stratified folds
    cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
    cv_scores = cross_val_score(classifier_pipeline, X_train, y_train, cv=cv, scoring='f1')
    logger.info(f"Cross-validation F1: {cv_scores.mean():.3f} (+/- {cv_scores.std() * 2:.3f})")

    # Test set evaluation
    y_pred = classifier_pipeline.predict(X_test)
    y_proba = classifier_pipeline.predict_proba(X_test)[:, 1]

    test_accuracy = accuracy_score(y_test, y_pred)
    test_f1 = f1_score(y_test, y_pred)
    test_auc = roc_auc_score(y_test, y_proba)

    logger.info(f"Test accuracy: {test_accuracy:.3f}")
    logger.info(f"Test F1 score: {test_f1:.3f}")
    logger.info(f"Test ROC-AUC: {test_auc:.3f}")

    # Detailed classification report
    report = classification_report(
        y_test, y_pred, target_names=['Legitimate', 'Misinformation'], output_dict=True
    )
    logger.info("Classification Report:")
    print(classification_report(y_test, y_pred, target_names=['Legitimate', 'Misinformation']))

    # Confusion matrix
    cm = confusion_matrix(y_test, y_pred)
    logger.info(f"Confusion Matrix:\n{cm}")

    # Save metrics to JSON
    metrics = {
        'timestamp': datetime.now().isoformat(),
        'dataset_size': len(df),
        'train_size': len(X_train),
        'test_size': len(X_test),
        'cv_f1_mean': round(cv_scores.mean(), 4),
        'cv_f1_std': round(cv_scores.std(), 4),
        'test_accuracy': round(test_accuracy, 4),
        'test_f1': round(test_f1, 4),
        'test_roc_auc': round(test_auc, 4),
        'classification_report': report,
        'confusion_matrix': cm.tolist()
    }

    metrics_path = Path(__file__).parent / 'models' / 'training_metrics.json'
    metrics_path.parent.mkdir(exist_ok=True)
    with open(metrics_path, 'w') as f:
        json.dump(metrics, f, indent=2)
    logger.info(f"Training metrics saved to {metrics_path}")

    # Sample predictions
    logger.info("\nSample Predictions:")
    test_samples = [
        "BREAKING: Government hiding shocking truth about vaccines - pharma exposed!",
        "Government announces new healthcare policy for rural India",
        "URGENT: Secret documents reveal conspiracy to overthrow democracy",
        "Research study published in Nature shows benefits of exercise",
        "VIRAL: This miracle cure will shock you - doctors HATE it!",
        "Parliament passes education reform bill with bipartisan support",
        "Hindu khatre mein hai - forward karo sabko ye video!",
        "According to ISRO chairman, Chandrayaan-3 data analysis is progressing well",
        "EXPOSED: EVM machines rigged - video proof showing hacking",
        "Fact-check: Viral claim about government banning cash is FALSE says RBI",
    ]

    for sample in test_samples:
        prediction = classifier_pipeline.predict([sample])[0]
        probability = classifier_pipeline.predict_proba([sample])[0]
        label = "FAKE" if prediction == 1 else "REAL"
        confidence = max(probability)
        fake_prob = probability[1] if len(probability) > 1 else 0.5
        logger.info(f"  [{label}] (conf={confidence:.3f}, fake_p={fake_prob:.3f}) "
                     f"'{sample[:60]}...'")

    return classifier_pipeline


def save_classifier(classifier, filename=None):
    """Save the trained classifier"""
    if filename is None:
        filename = str(Path(__file__).parent / 'models' / 'advanced_misinformation_classifier.pkl')
    Path(filename).parent.mkdir(exist_ok=True)
    with open(filename, 'wb') as f:
        pickle.dump(classifier, f)
    logger.info(f"Classifier saved to {filename}")


class _CustomUnpickler(pickle.Unpickler):
    """Custom unpickler that resolves classes saved under __main__ back to this module.

    When advanced_ml_classifier.py is run directly (python advanced_ml_classifier.py),
    pickle stores custom classes as '__main__.LinguisticFeatureExtractor' etc.
    When those pickled models are later loaded from a different entry-point
    (e.g. main_application.py), Python tries to find the class in __main__ and fails.
    This unpickler transparently redirects those lookups back to this module.
    """
    def find_class(self, module, name):
        if module == '__main__':
            import advanced_ml_classifier
            if hasattr(advanced_ml_classifier, name):
                return getattr(advanced_ml_classifier, name)
        return super().find_class(module, name)


def load_classifier(filename=None):
    """Load a trained classifier"""
    if filename is None:
        filename = str(Path(__file__).parent / 'models' / 'advanced_misinformation_classifier.pkl')
    try:
        with open(filename, 'rb') as f:
            classifier = _CustomUnpickler(f).load()
        logger.info(f"Classifier loaded from {filename}")
        return classifier
    except FileNotFoundError:
        logger.warning(f"Classifier file {filename} not found, building new one...")
        classifier = build_advanced_classifier()
        save_classifier(classifier, filename)
        return classifier


if __name__ == "__main__":
    print("🧠 Building Advanced ML Classifier for Misinformation Detection")
    print("=" * 70)

    # Build and train classifier
    classifier = build_advanced_classifier()

    # Save classifier
    save_classifier(classifier)

    print("\n✅ Advanced classifier built and saved successfully!")
    print("📊 Ready for integration with real-time system")