explain.py

import numpy as np
from joblib import Parallel, delayed, parallel_backend

class Explain:
    """
    Computes feature importance using various interpretability strategies.
    """

    def __init__(self, importance_type='default'):
        self.importance_type = importance_type

    def get_importance(self, predictor, val_df, label_column='class', n_repeats=1):
        if self.importance_type == 'default':
            return self._default_importance(predictor)
        elif self.importance_type == 'permutation':
            return self._permutation_importance(predictor, val_df, label_column, n_repeats)
        else:
            raise ValueError(f"Unknown importance_type: {self.importance_type}")

    def _default_importance(self, predictor):
        print(f"[Explain] Default importance from model {predictor.model_best}")
        best_model = predictor._trainer.load_model(predictor.model_best).model
        feature_names = predictor.feature_metadata.get_features()
        importances = best_model.feature_importances_
        return dict(zip(feature_names, importances))

    def _permutation_importance(self, predictor, val_df, label_column, n_repeats):
        print("[Explain] Permutation importance")
        np.random.seed(42)  # For reproducibility

        base_feature = predictor._learner.feature_generator.feature_metadata_in.get_features()
        base_acc = predictor.evaluate(val_df, silent=False)['accuracy']
        
        importance = {}
        for feature in base_feature:
            drops = []
            for _ in range(n_repeats):
                tmp = val_df.copy()
                tmp[feature] = np.random.permutation(tmp[feature])
                perm_acc = predictor.evaluate(tmp, silent=True)['accuracy']
                drops.append(base_acc - perm_acc)
            importance[feature] = np.mean(drops)
        return importance

    def _get_feature_impact(self, predictor, df, base_acc):
        perm_acc = predictor.evaluate(df, silent=True)['accuracy']
        return base_acc - perm_acc