File size: 13,601 Bytes
59f824e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Core Gradient Boosting classifier for deception detection.



This module implements a robust, production-oriented classifier wrapper that

standardizes features, handles missing values, performs probability calibration,

and persists all metadata required for reproducible inference.

"""

from __future__ import annotations

from dataclasses import asdict
from typing import Any, Dict, List, Optional, Tuple

import numpy as np
import pandas as pd
from sklearn.calibration import CalibratedClassifierCV
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.metrics import accuracy_score, log_loss, roc_auc_score
from sklearn.model_selection import cross_val_score
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler

from src.utils.config import ModelConfig
from src.utils.io import load_model as io_load_model
from src.utils.io import save_model as io_save_model


class DeceptionClassifier:
    """Gradient Boosting-based deception classifier with probability calibration.



    Uses `sklearn` GradientBoostingClassifier as the base estimator and wraps it

    with calibration for better probabilistic interpretation.

    """

    def __init__(self, config: Optional[ModelConfig] = None):
        """Initialize classifier with model config parameters.



        Args:

            config: Optional model configuration. If not provided, defaults are used.

        """

        self.config: ModelConfig = config or ModelConfig()
        self.model = GradientBoostingClassifier(
            n_estimators=self.config.n_estimators,
            learning_rate=self.config.learning_rate,
            max_depth=self.config.max_depth,
            min_samples_split=self.config.min_samples_split,
            min_samples_leaf=self.config.min_samples_leaf,
            subsample=self.config.subsample,
            random_state=self.config.random_state,
        )
        self.scaler = StandardScaler()

        calibration_method = getattr(self.config, "calibration_method", "sigmoid")
        self.calibrated_model = CalibratedClassifierCV(
            estimator=GradientBoostingClassifier(
                n_estimators=self.config.n_estimators,
                learning_rate=self.config.learning_rate,
                max_depth=self.config.max_depth,
                min_samples_split=self.config.min_samples_split,
                min_samples_leaf=self.config.min_samples_leaf,
                subsample=self.config.subsample,
                random_state=self.config.random_state,
            ),
            method=calibration_method,
            cv=min(max(2, getattr(self.config, "cross_validation_folds", 5)), 5),
        )

        self.feature_names_: Optional[List[str]] = None
        self.is_fitted_: bool = False
        self.training_metadata_: Dict[str, Any] = {}

    def prepare_features(

        self,

        features_df: pd.DataFrame,

        feature_columns: Optional[List[str]] = None,

    ) -> Tuple[np.ndarray, List[str]]:
        """Prepare feature matrix from a feature dataframe.



        Args:

            features_df: DataFrame with extracted features.

            feature_columns: Optional explicit feature column list.



        Returns:

            Tuple[np.ndarray, List[str]]: Feature matrix and ordered feature names.

        """

        if features_df is None or features_df.empty:
            return np.empty((0, 0), dtype=np.float64), []

        excluded_columns = {
            "frame_idx",
            "timestamp",
            "segment_idx",
            "start_time",
            "end_time",
            "label",
            "subject_id",
        }

        if feature_columns is None:
            if self.is_fitted_ and self.feature_names_ is not None:
                feature_columns = self.feature_names_
            else:
                candidate_cols = [c for c in features_df.columns if c not in excluded_columns]
                # Research rationale: constraining to numeric-only columns prevents accidental
                # leakage of identifiers and ensures stable scaling behavior.
                feature_columns = [
                    c for c in candidate_cols if pd.api.types.is_numeric_dtype(features_df[c])
                ]

        if not feature_columns:
            return np.empty((len(features_df), 0), dtype=np.float64), []

        missing_cols = set(feature_columns) - set(features_df.columns)
        if missing_cols:
            for col in missing_cols:
                features_df[col] = 0.0

        selected = features_df[feature_columns].copy()
        for col in selected.columns:
            selected[col] = pd.to_numeric(selected[col], errors="coerce")

        selected = selected.replace([np.inf, -np.inf], np.nan)
        medians = selected.median(numeric_only=True)
        selected = selected.fillna(medians)
        selected = selected.fillna(0.0)

        # Clip to robust quantiles to reduce effect of extreme outliers.
        lower = selected.quantile(0.01)
        upper = selected.quantile(0.99)
        selected = selected.clip(lower=lower, upper=upper, axis=1)

        return selected.to_numpy(dtype=np.float64), list(selected.columns)

    def train(self, X: np.ndarray, y: np.ndarray, feature_names: List[str]) -> Dict[str, float]:
        """Train the gradient boosting classifier.



        Args:

            X: Feature matrix of shape (n_samples, n_features).

            y: Binary labels (0=truthful, 1=deceptive).

            feature_names: Ordered feature names.



        Returns:

            Dict[str, float]: Training metrics and dataset summary.

        """

        X = np.asarray(X, dtype=np.float64)
        y = np.asarray(y).astype(int)

        if X.ndim != 2:
            raise ValueError("X must be a 2D array.")
        if y.ndim != 1:
            raise ValueError("y must be a 1D array.")
        if len(X) != len(y):
            raise ValueError("X and y must have the same number of samples.")
        if len(y) == 0:
            raise ValueError("Cannot train with zero samples.")

        self.feature_names_ = list(feature_names)

        X_scaled = self.scaler.fit_transform(X)

        unique_classes, counts = np.unique(y, return_counts=True)
        class_distribution = {int(k): int(v) for k, v in zip(unique_classes, counts)}

        # Research rationale: inverse-frequency weighting mitigates skewed priors and
        # helps gradient boosting focus on under-represented deceptive/truthful samples.
        class_weights = {
            cls: len(y) / (len(unique_classes) * count)
            for cls, count in zip(unique_classes, counts)
            if count > 0
        }
        sample_weight = np.array([class_weights.get(label, 1.0) for label in y], dtype=np.float64)

        self.model.fit(X_scaled, y, sample_weight=sample_weight)

        can_calibrate = len(unique_classes) > 1 and len(y) >= max(10, len(unique_classes) * 3)
        if can_calibrate:
            self.calibrated_model.fit(X_scaled, y, sample_weight=sample_weight)

        # Research rationale: setting is_fitted_ prior to metric computation allows
        # reuse of internal predict/predict_proba methods for training performance tracking.
        self.is_fitted_ = True

        train_preds = self.predict(X, check_fitted=False)
        train_probs = self.predict_proba(X, check_fitted=False)

        metrics: Dict[str, float] = {
            "train_accuracy": float(accuracy_score(y, train_preds)),
            "n_samples": float(len(y)),
            "n_features": float(X.shape[1]),
            "class_distribution": class_distribution,  # type: ignore[assignment]
        }

        try:
            metrics["train_log_loss"] = float(log_loss(y, train_probs, labels=[0, 1]))
        except ValueError:
            metrics["train_log_loss"] = float("nan")

        if len(unique_classes) > 1:
            try:
                metrics["train_auc"] = float(roc_auc_score(y, train_probs[:, 1]))
            except ValueError:
                metrics["train_auc"] = float("nan")
        else:
            metrics["train_auc"] = float("nan")

        cv_folds = min(getattr(self.config, "cross_validation_folds", 5), len(y))
        if cv_folds >= 2 and len(unique_classes) > 1:
            pipeline = make_pipeline(
                StandardScaler(),
                GradientBoostingClassifier(
                    n_estimators=self.config.n_estimators,
                    learning_rate=self.config.learning_rate,
                    max_depth=self.config.max_depth,
                    min_samples_split=self.config.min_samples_split,
                    min_samples_leaf=self.config.min_samples_leaf,
                    subsample=self.config.subsample,
                    random_state=self.config.random_state,
                ),
            )
            cv_scores = cross_val_score(pipeline, X, y, cv=cv_folds, scoring="accuracy")
            metrics["cv_accuracy_mean"] = float(np.mean(cv_scores))
            metrics["cv_accuracy_std"] = float(np.std(cv_scores))
        else:
            metrics["cv_accuracy_mean"] = float("nan")
            metrics["cv_accuracy_std"] = float("nan")

        self.training_metadata_ = metrics
        return metrics

    def predict(self, X: np.ndarray, check_fitted: bool = True) -> np.ndarray:
        """Predict binary deception labels for input features.

        

        Args:

            X: Input feature matrix.

            check_fitted: Whether to verify that the model is fitted.

        """

        if check_fitted and not self.is_fitted_:
            raise ValueError("Model is not trained. Call train() first.")

        X_scaled = self.scaler.transform(np.asarray(X, dtype=np.float64))
        if hasattr(self.calibrated_model, "calibrated_classifiers_"):
            return self.calibrated_model.predict(X_scaled)
        return self.model.predict(X_scaled)

    def predict_proba(self, X: np.ndarray, check_fitted: bool = True) -> np.ndarray:
        """Predict deception probabilities for input features.



        Args:

            X: Input feature matrix.

            check_fitted: Whether to verify that the model is fitted.

            

        Returns:

            np.ndarray: Probability matrix [truthful_prob, deceptive_prob].

        """

        if check_fitted and not self.is_fitted_:
            raise ValueError("Model is not trained. Call train() first.")

        X_scaled = self.scaler.transform(np.asarray(X, dtype=np.float64))
        if hasattr(self.calibrated_model, "calibrated_classifiers_"):
            return self.calibrated_model.predict_proba(X_scaled)
        return self.model.predict_proba(X_scaled)

    def predict_deception_score(self, X: np.ndarray, check_fitted: bool = True) -> np.ndarray:
        """Predict scalar deception score in [0, 1] for each sample."""

        proba = self.predict_proba(X, check_fitted=check_fitted)
        return proba[:, 1]

    def get_feature_importance(self) -> pd.DataFrame:
        """Get sorted feature importance values from the trained base model."""

        if not self.is_fitted_:
            raise ValueError("Model is not trained. Call train() first.")

        if self.feature_names_ is None:
            raise ValueError("Feature names are unavailable.")

        importances = getattr(self.model, "feature_importances_", None)
        if importances is None:
            raise ValueError("Model does not expose feature importances.")

        importance_df = pd.DataFrame(
            {"feature": self.feature_names_, "importance": np.asarray(importances, dtype=np.float64)}
        )
        return importance_df.sort_values("importance", ascending=False).reset_index(drop=True)

    def save(self, path: str) -> None:
        """Save model, scaler, calibration wrapper, and metadata using joblib."""

        payload = {
            "model": self.model,
            "calibrated_model": self.calibrated_model,
            "scaler": self.scaler,
            "feature_names": self.feature_names_,
            "is_fitted": self.is_fitted_,
            "config": asdict(self.config),
            "training_metadata": self.training_metadata_,
        }
        io_save_model(payload, path)

    def load(self, path: str) -> None:
        """Load model, scaler, calibration wrapper, and metadata from file."""

        payload = io_load_model(path)
        self.model = payload["model"]
        self.calibrated_model = payload.get("calibrated_model", self.calibrated_model)
        self.scaler = payload["scaler"]
        self.feature_names_ = payload.get("feature_names")
        self.is_fitted_ = bool(payload.get("is_fitted", True))
        self.training_metadata_ = dict(payload.get("training_metadata", {}))

        config_dict = payload.get("config", {})
        if isinstance(config_dict, dict):
            self.config = ModelConfig(**config_dict)

    def get_model_info(self) -> Dict[str, Any]:
        """Return model metadata and configuration details."""

        return {
            "is_fitted": self.is_fitted_,
            "n_features": 0 if self.feature_names_ is None else len(self.feature_names_),
            "feature_names": self.feature_names_ or [],
            "config": asdict(self.config),
            "training_metadata": self.training_metadata_,
        }