fix: 🐛 fix shap imp calculation for multiclass

src/arfs/feature_selection/variable_importance.py

@@ -398,15 +398,21 @@ def _compute_varimp_lgb(
                 # X_SHAP_values (array-like of shape = [n_samples, n_features + 1]
                 # or shape = [n_samples, (n_features + 1) * n_classes])
                 # index starts from 0
-                n_feat = gbm_model.valid_features.shape[1]
+                n_features_plus_bias = gbm_model.valid_features.shape[1] + 1
+                n_samples = gbm_model.valid_features.shape[0]
                 y_freq_table = pd.Series(y.fillna(0)).value_counts(normalize=True)
                 n_classes = y_freq_table.size
-                shap_matrix = np.delete(
-                    shap_matrix,
-                    list(range(n_feat, (n_feat + 1) * n_classes, n_feat + 1)),
-                    axis=1,
-                )
-                shap_imp = np.mean(np.abs(shap_matrix[:, :-1]), axis=0)
+
+                # Reshape the array to [n_samples, n_features + 1, n_classes]
+                reshaped_values = shap_matrix.reshape(n_samples, n_classes, n_features_plus_bias)
+
+                # Since we need (n_samples, n_features + 1, n_classes), transpose the second and third dimensions
+                reshaped_values = reshaped_values.transpose(0, 2, 1)
+                reshaped_values = reshaped_values[:, :-1, :]
+                reshaped_values.shape
+                # Sum the contributions for each class ignoring the bias term
+                # average on all the samples
+                shap_imp = np.abs(reshaped_values).sum(axis=-1).mean(axis=0)
             else:
                 # for binary, only one class is returned, for regression a single column added as well
                 shap_imp = np.mean(np.abs(shap_matrix[:, :-1]), axis=0)