scripts/models.py

import numpy as np

class SOTAStackingDetector:
    """State-of-the-Art Stacked Ensemble Classifier for AI text detection.
    Combines stylistic baselines, tree-based stylometrics, and vocabulary-based n-grams
    using a Meta-Classifier to achieve robust generalization.
    """
    def __init__(self, lr_sty, xgb_sty, lr_ng, meta_model, num_sty_features):
        self.lr_sty = lr_sty
        self.xgb_sty = xgb_sty
        self.lr_ng = lr_ng
        self.meta_model = meta_model
        self.num_sty_features = num_sty_features
        # Compatibility fields for SHAP / Explainability coefficients
        self.coef_ = lr_sty.coef_
        self.intercept_ = lr_sty.intercept_
        
    def _extract_base_features(self, X_hybrid_scaled):
        X_sty_scaled = X_hybrid_scaled[:, :self.num_sty_features]
        X_ng_scaled = X_hybrid_scaled[:, self.num_sty_features:]
        return X_sty_scaled, X_ng_scaled
        
    def predict_proba(self, X_hybrid_scaled):
        X_sty_scaled, X_ng_scaled = self._extract_base_features(X_hybrid_scaled)
        
        # Get base model probabilities
        p_lr_sty = self.lr_sty.predict_proba(X_sty_scaled)[:, 1]
        p_xgb_sty = self.xgb_sty.predict_proba(X_sty_scaled)[:, 1]
        p_lr_ng = self.lr_ng.predict_proba(X_ng_scaled)[:, 1]
        
        # Construct meta features
        X_meta = np.column_stack([p_lr_sty, p_xgb_sty, p_lr_ng])
        return self.meta_model.predict_proba(X_meta)
        
    def predict(self, X_hybrid_scaled):
        prob = self.predict_proba(X_hybrid_scaled)[:, 1]
        return (prob >= 0.5).astype(int)