File size: 11,595 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
"""Ablation study for feature group importance in deception detection.



This module provides systematic removal of feature groups (facial, audio,

temporal, cross‑modal) to quantify their contribution to model performance.

"""

from __future__ import annotations

import logging
from typing import Any, Dict, List, Optional, Tuple, Union

import numpy as np
import pandas as pd
from sklearn.base import BaseEstimator, clone

from .cross_validation import CrossValidator
from .metrics import DeceptionMetrics


logger = logging.getLogger(__name__)


class AblationStudy:
    """Conduct ablation studies by removing feature groups.



    This class evaluates the contribution of different feature groups

    (facial, audio, temporal, cross‑modal) to overall model performance.



    Attributes:

        feature_names: List of all feature names in the dataset.

        feature_groups: Dictionary mapping group names to lists of feature indices or names.

        random_state: Random seed for reproducibility.

        n_splits: Number of cross‑validation folds.

        metrics: Instance of DeceptionMetrics for evaluation.

    """

    def __init__(

        self,

        feature_names: List[str],

        feature_groups: Dict[str, List[str]],

        random_state: int = 42,

        n_splits: int = 5,

        metrics: Optional[DeceptionMetrics] = None,

    ) -> None:
        """Initialize ablation study.



        Args:

            feature_names: List of all feature names in the dataset.

            feature_groups: Dictionary mapping group names to lists of feature names.

                Expected keys: 'facial', 'audio', 'temporal', 'cross_modal'.

            random_state: Random seed for reproducibility.

            n_splits: Number of cross‑validation folds.

            metrics: Optional DeceptionMetrics instance.

        """
        self.feature_names = feature_names
        self.feature_groups = feature_groups
        self.random_state = random_state
        self.n_splits = n_splits
        self.metrics = metrics or DeceptionMetrics(random_state=random_state)

        # Validate that all group features are in feature_names
        for group_name, features in feature_groups.items():
            for f in features:
                if f not in feature_names:
                    logger.warning(
                        f"Feature '{f}' from group '{group_name}' not found in feature_names"
                    )

        # Build index mapping
        self.feature_to_index = {name: i for i, name in enumerate(feature_names)}
        self.group_indices = {}
        for group_name, features in feature_groups.items():
            indices = []
            for f in features:
                if f in self.feature_to_index:
                    indices.append(self.feature_to_index[f])
            self.group_indices[group_name] = indices

    def _select_features(

        self,

        X: Union[np.ndarray, pd.DataFrame],

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

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

    ) -> Union[np.ndarray, pd.DataFrame]:
        """Select features based on included/excluded groups.



        Args:

            X: Original feature matrix (n_samples × n_features).

            include_groups: List of group names to keep (all others removed).

            exclude_groups: List of group names to remove (all others kept).



        Returns:

            Feature matrix with selected columns.

        """
        if include_groups is not None and exclude_groups is not None:
            raise ValueError("Cannot specify both include_groups and exclude_groups")

        if isinstance(X, pd.DataFrame):
            columns = X.columns.tolist()
        else:
            columns = list(range(X.shape[1]))

        # Determine which indices to keep
        if include_groups is not None:
            keep_indices = set()
            for group in include_groups:
                keep_indices.update(self.group_indices.get(group, []))
        elif exclude_groups is not None:
            all_indices = set(range(len(self.feature_names)))
            remove_indices = set()
            for group in exclude_groups:
                remove_indices.update(self.group_indices.get(group, []))
            keep_indices = all_indices - remove_indices
        else:
            # Keep all features
            keep_indices = set(range(len(self.feature_names)))

        # Convert to sorted list
        keep_indices = sorted(keep_indices)

        if isinstance(X, pd.DataFrame):
            selected_cols = [columns[i] for i in keep_indices]
            return X[selected_cols]
        else:
            return X[:, keep_indices]

    def run_ablation(

        self,

        classifier_class: type,

        X: Union[np.ndarray, pd.DataFrame],

        y: Union[np.ndarray, List[int]],

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

        feature_groups: Optional[Dict[str, List[str]]] = None,

        n_splits: Optional[int] = None,

        subject_ids: Optional[Union[np.ndarray, List[str]]] = None,

        cv_method: str = "stratified",

    ) -> Dict[str, Dict[str, Any]]:
        """Run ablation study across feature groups.



        Args:

            classifier_class: A scikit‑learn classifier class (not instance).

            X: Feature matrix.

            y: Target labels.

            feature_names: Override stored feature names.

            feature_groups: Override stored feature groups.

            n_splits: Override default number of folds.

            subject_ids: Optional subject IDs for subject‑aware CV.

            cv_method: Cross‑validation method ('stratified', 'subject_aware', 'loso').



        Returns:

            Dictionary with keys:

                - full_model: metrics with all features

                - minus_facial: metrics without facial features

                - minus_audio: metrics without audio features

                - minus_temporal: metrics without temporal features

                - minus_cross_modal: metrics without cross‑modal features

                - facial_only: metrics with only facial features

                - audio_only: metrics with only audio features

                - temporal_only: metrics with only temporal features

            Each value is a dict with 'mean_metrics', 'std_metrics', 'per_fold_metrics'.

        """
        if feature_names is not None:
            self.feature_names = feature_names
        if feature_groups is not None:
            self.feature_groups = feature_groups
            self.group_indices = {}
            for group_name, features in feature_groups.items():
                indices = []
                for f in features:
                    if f in self.feature_to_index:
                        indices.append(self.feature_to_index[f])
                self.group_indices[group_name] = indices

        n_splits = n_splits or self.n_splits
        cv = CrossValidator(
            random_state=self.random_state,
            n_splits=n_splits,
            metrics=self.metrics,
        )

        # Define ablation conditions
        conditions = {
            "full_model": {"include_groups": None, "exclude_groups": None},
            "minus_facial": {"exclude_groups": ["facial"]},
            "minus_audio": {"exclude_groups": ["audio"]},
            "minus_temporal": {"exclude_groups": ["temporal"]},
            "minus_cross_modal": {"exclude_groups": ["cross_modal"]},
            "facial_only": {"include_groups": ["facial"]},
            "audio_only": {"include_groups": ["audio"]},
            "temporal_only": {"include_groups": ["temporal"]},
        }

        results = {}

        for cond_name, cond_args in conditions.items():
            logger.info(f"Running ablation condition: {cond_name}")

            # Select features
            X_cond = self._select_features(
                X,
                include_groups=cond_args.get("include_groups"),
                exclude_groups=cond_args.get("exclude_groups"),
            )

            # Skip if no features remain
            if X_cond.shape[1] == 0:
                logger.warning(f"No features left for condition {cond_name}, skipping")
                results[cond_name] = {
                    "mean_metrics": {},
                    "std_metrics": {},
                    "per_fold_metrics": [],
                }
                continue

            # Run cross‑validation
            classifier = classifier_class()
            cv_result = cv.run_cross_validation(
                classifier,
                X_cond,
                y,
                n_splits=n_splits,
                subject_ids=subject_ids,
                method=cv_method,
                return_predictions=False,
            )

            results[cond_name] = {
                "mean_metrics": cv_result["mean_metrics"],
                "std_metrics": cv_result["std_metrics"],
                "per_fold_metrics": cv_result["per_fold_metrics"],
                "n_features": X_cond.shape[1],
            }

        logger.info("Ablation study completed.")
        return results

    def compute_feature_importance_ranking(

        self,

        classifier: BaseEstimator,

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

    ) -> List[Tuple[str, float]]:
        """Compute feature importance ranking from a trained classifier.



        Supports classifiers with `feature_importances_` attribute (tree‑based)

        or `coef_` attribute (linear models).



        Args:

            classifier: Trained scikit‑learn classifier.

            feature_names: Override stored feature names.



        Returns:

            List of (feature_name, importance_score) sorted descending.

        """
        if feature_names is None:
            feature_names = self.feature_names

        if hasattr(classifier, "feature_importances_"):
            importances = classifier.feature_importances_
        elif hasattr(classifier, "coef_"):
            # For linear models, take absolute coefficients averaged across classes
            coef = classifier.coef_
            if coef.ndim == 2:
                importances = np.mean(np.abs(coef), axis=0)
            else:
                importances = np.abs(coef)
        else:
            logger.warning(
                "Classifier has no feature_importances_ or coef_; returning uniform importances"
            )
            importances = np.ones(len(feature_names))

        # Ensure length matches
        if len(importances) != len(feature_names):
            logger.error(
                f"Importance length ({len(importances)}) != feature count ({len(feature_names)})"
            )
            # Truncate or pad
            if len(importances) > len(feature_names):
                importances = importances[: len(feature_names)]
            else:
                importances = np.pad(
                    importances,
                    (0, len(feature_names) - len(importances)),
                    mode="constant",
                    constant_values=0.0,
                )

        ranked = sorted(
            zip(feature_names, importances),
            key=lambda x: x[1],
            reverse=True,
        )
        logger.info(f"Computed feature importance ranking for {len(ranked)} features")
        return ranked