models.py

import pandas as pd
import numpy as np
from sklearn.naive_bayes import MultinomialNB, ComplementNB
from sklearn.linear_model import LogisticRegression
from sklearn.svm import LinearSVC
from sklearn.calibration import CalibratedClassifierCV
from sklearn.ensemble import (
    VotingClassifier, RandomForestClassifier, StackingClassifier,
    GradientBoostingClassifier, HistGradientBoostingClassifier
)
from sklearn.feature_selection import SelectKBest, f_classif
from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import MinMaxScaler
from sklearn.feature_extraction.text import TfidfVectorizer
from dl_model import DeepNewsClassifier  # §3.3.2 Deep Learning Module


class FakeNewsModels:
    def __init__(self):
        # All hand-crafted numeric columns (original + new deception features)
        num_cols = [
            # Original features
            'lexical_density', 'capital_ratio', 'noun_ratio', 'speaker_reliability',
            'sentiment_score', 'complexity_score', 'subjectivity_score',
            'verb_ratio', 'adj_ratio', 'adv_ratio', 'punctuation_aggression',
            'entity_density', 'formal_cadence', 'official_marker',
            # Hedging & Certainty
            'hedge_ratio', 'certainty_ratio', 'hedge_certainty_diff',
            # Emotional
            'emotional_intensity', 'sentiment_extremity',
            'negativity_score', 'positivity_score',
            # Numerical
            'number_density', 'has_percentage', 'has_dollar_amount', 'number_count',
            # Source / Attribution
            'attribution_ratio', 'has_source_citation',
            # Readability
            'ari_score', 'gunning_fog', 'avg_sentence_length',
            # Pronoun & Modality
            'first_person_ratio', 'modal_ratio', 'negation_ratio',
            # Comparison
            'comparative_ratio', 'superlative_ratio',
            # Voice
            'passive_ratio',
            # Urgency / Style
            'caps_word_ratio', 'ellipsis_count',
            'exclamation_marks', 'question_marks', 'quotes_count',
            'total_words', 'avg_word_length',
            # Social context history
            'false_history_ratio', 'true_history_ratio', 'history_volume',
            'is_republican', 'is_democrat',
        ]

        preprocessor = ColumnTransformer(
            transformers=[
                ('text', TfidfVectorizer(
                    ngram_range=(1, 3),
                    max_features=12000,
                    min_df=3,
                    max_df=0.75,
                    sublinear_tf=True,
                    stop_words='english'
                ), 'statement'),
                ('meta', MinMaxScaler(), num_cols),
            ],
            remainder='drop'
        )

        # Feature selection: keep top 5000 features
        selector = SelectKBest(f_classif, k=5000)

        # Individual classifiers — tuned for better performance and calibrated uncertainty
        nb_clf = CalibratedClassifierCV(ComplementNB(alpha=0.1))
        lr_clf = LogisticRegression(
            C=0.8, max_iter=3000, solver='lbfgs',
            class_weight='balanced', l1_ratio=0
        )
        svm_clf = CalibratedClassifierCV(
            LinearSVC(C=0.5, max_iter=5000, class_weight='balanced', dual='auto')
        )
        rf_clf = CalibratedClassifierCV(
            RandomForestClassifier(
                n_estimators=600, max_depth=22, min_samples_leaf=3,
                class_weight='balanced', random_state=42, n_jobs=-1
            )
        )

        self.models = {
            'nb':  Pipeline([('pre', preprocessor), ('sel', selector), ('clf', nb_clf)]),
            'lr':  Pipeline([('pre', preprocessor), ('sel', selector), ('clf', lr_clf)]),
            'svm': Pipeline([('pre', preprocessor), ('sel', selector), ('clf', svm_clf)]),
            'rf':  Pipeline([('pre', preprocessor), ('sel', selector), ('clf', rf_clf)]),
            # §3.3.2 Deep Learning Module — BiLSTM-approximating dual encoder
            'dl':  DeepNewsClassifier(),
        }

        # STACKING ENSEMBLE with LogisticRegression meta-learner (§3.3)
        # Includes Deep Learning module as 5th base estimator
        self.models['ensemble'] = StackingClassifier(
            estimators=[
                ('nb',  self.models['nb']),
                ('lr',  self.models['lr']),
                ('svm', self.models['svm']),
                ('rf',  self.models['rf']),
                ('dl',  self.models['dl']),  # §3.3.2
            ],
            final_estimator=LogisticRegression(
                C=0.5, class_weight='balanced', max_iter=1000
            ),
            stack_method='predict_proba',
            cv=5,
            n_jobs=-1
        )

    def get_model(self, name):
        return self.models.get(name)

    def get_all_models(self):
        return self.models