Expand source code
# generic imports
import numpy as np
# imports from imodels
from imodels.tree.rf_plus.feature_importance.ppms.ppms import MDIPlusGenericRegressorPPM, MDIPlusGenericClassifierPPM
class LMDIPlus():
"""
Local MDI+ (LMDI+) Explainer for tree-based models.
Parameters:
----------
rf_plus_model : RFPlusModel
A trained RF+ model.
evaluate_on : str
Specifies which samples to use when computing local importances:
- 'all' (default): evaluate on all samples.
- 'oob': only on out-of-bag samples per tree.
- 'inbag': only on in-bag samples per tree.
"""
def __init__(self, rf_plus_model, evaluate_on = 'all'):
self.rf_plus_model = rf_plus_model
self.mode = 'only_k'
self.oob_indices = self.rf_plus_model._oob_indices
self.evaluate_on = evaluate_on
if self.rf_plus_model._task == "classification":
self.tree_explainers = [MDIPlusGenericClassifierPPM(rf_plus_model.estimators_[i])
for i in range(len(rf_plus_model.estimators_))]
else:
self.tree_explainers = [MDIPlusGenericRegressorPPM(rf_plus_model.estimators_[i])
for i in range(len(rf_plus_model.estimators_))]
def get_lmdi_plus_scores(self, X, y = None, njobs = 1, ranking = False):
"""
Compute LMDI+ scores for each sample in X.
If `y` is provided, the evaluation is assumed to be on the training set, and if `y` is None,
it assumes LFI computation is being performed on unseen test data.
Parameters:
----------
X : np.ndarray
Input feature matrix of shape (n_samples, n_features).
y : np.ndarray or None
Target values for the training set or None for test data.
njobs : int
Number of parallel jobs to use for prediction (if supported).
ranking : bool
If True, converts the LFI scores to feature rankings per sample.
Returns:
-------
local_feature_importances : np.ndarray
Matrix of shape (n_samples, n_features) containing the averaged LMDI+ scores across trees.
"""
local_feature_importances = np.full((X.shape[0], X.shape[1],len(self.tree_explainers)), np.nan)
if y is None:
evaluate_on = None
else:
evaluate_on = self.evaluate_on
lfi_scores = self._get_LFI_subtract_intercept(X, njobs)
for i in range(lfi_scores.shape[-1]):
ith_tree_scores = lfi_scores[:, :, i]
oob_indices = np.unique(self.oob_indices[i])
if evaluate_on == 'oob':
local_feature_importances[oob_indices, :, i] = \
ith_tree_scores[oob_indices, :]
elif evaluate_on == 'inbag':
inbag_indices = np.arange(X.shape[0])
inbag_indices = np.setdiff1d(inbag_indices, oob_indices)
local_feature_importances[inbag_indices, :, i] = \
ith_tree_scores[inbag_indices, :]
else:
local_feature_importances[:, :, i] = ith_tree_scores
if ranking:
local_feature_importances = np.abs(local_feature_importances)
rank_matrix = np.zeros_like(local_feature_importances)
for i in range(local_feature_importances.shape[-1]):
lfi_treei = local_feature_importances[:,:,i]
indices_of_zero_columns = np.where(np.all(lfi_treei==0, axis=0))[0]
lfi_treei[:, indices_of_zero_columns] = -1
ranks = np.argsort(np.argsort(lfi_treei, kind="stable"), kind = "stable")
ranks = np.array(ranks, dtype=np.float32)
ranks[:, indices_of_zero_columns] = np.nan
rank_matrix[:,:,i] = ranks
local_feature_importances = rank_matrix
local_feature_importances = np.nanmean(local_feature_importances, axis=-1)
local_feature_importances[np.isnan(local_feature_importances)] = 0
return local_feature_importances
def _get_LFI_subtract_intercept(self, X, njobs):
"""
Compute per-tree LMDI+ scores for each sample in X.
Parameters:
----------
X : np.ndarray
Input feature matrix of shape (n_samples, n_features).
njobs : int
Number of parallel jobs to use for prediction (if supported).
Returns:
-------
LFIs : np.ndarray
Matrix of shape (n_samples, n_features, n_trees) containing the per-tree
local feature importance scores.
"""
LFIs = np.zeros((X.shape[0],X.shape[1],len(self.tree_explainers)))
for i, tree_explainer in enumerate(self.tree_explainers):
blocked_data_ith_tree = self.rf_plus_model.transformers_[i].transform(X)
if self.rf_plus_model._task == "classification":
ith_partial_preds = tree_explainer.predict_partial_subtract_intercept(blocked_data_ith_tree, njobs=njobs)
else:
ith_partial_preds = tree_explainer.predict_partial_subtract_intercept(blocked_data_ith_tree, njobs=njobs)
ith_partial_preds = np.array([ith_partial_preds[j] for j in range(X.shape[1])]).T
LFIs[:,:,i] = ith_partial_preds
return LFIsClasses
class LMDIPlus (rf_plus_model, evaluate_on='all')-
Local MDI+ (LMDI+) Explainer for tree-based models.
Parameters:
rf_plus_model : RFPlusModel A trained RF+ model. evaluate_on : str Specifies which samples to use when computing local importances: - 'all' (default): evaluate on all samples. - 'oob': only on out-of-bag samples per tree. - 'inbag': only on in-bag samples per tree.
Expand source code
class LMDIPlus(): """ Local MDI+ (LMDI+) Explainer for tree-based models. Parameters: ---------- rf_plus_model : RFPlusModel A trained RF+ model. evaluate_on : str Specifies which samples to use when computing local importances: - 'all' (default): evaluate on all samples. - 'oob': only on out-of-bag samples per tree. - 'inbag': only on in-bag samples per tree. """ def __init__(self, rf_plus_model, evaluate_on = 'all'): self.rf_plus_model = rf_plus_model self.mode = 'only_k' self.oob_indices = self.rf_plus_model._oob_indices self.evaluate_on = evaluate_on if self.rf_plus_model._task == "classification": self.tree_explainers = [MDIPlusGenericClassifierPPM(rf_plus_model.estimators_[i]) for i in range(len(rf_plus_model.estimators_))] else: self.tree_explainers = [MDIPlusGenericRegressorPPM(rf_plus_model.estimators_[i]) for i in range(len(rf_plus_model.estimators_))] def get_lmdi_plus_scores(self, X, y = None, njobs = 1, ranking = False): """ Compute LMDI+ scores for each sample in X. If `y` is provided, the evaluation is assumed to be on the training set, and if `y` is None, it assumes LFI computation is being performed on unseen test data. Parameters: ---------- X : np.ndarray Input feature matrix of shape (n_samples, n_features). y : np.ndarray or None Target values for the training set or None for test data. njobs : int Number of parallel jobs to use for prediction (if supported). ranking : bool If True, converts the LFI scores to feature rankings per sample. Returns: ------- local_feature_importances : np.ndarray Matrix of shape (n_samples, n_features) containing the averaged LMDI+ scores across trees. """ local_feature_importances = np.full((X.shape[0], X.shape[1],len(self.tree_explainers)), np.nan) if y is None: evaluate_on = None else: evaluate_on = self.evaluate_on lfi_scores = self._get_LFI_subtract_intercept(X, njobs) for i in range(lfi_scores.shape[-1]): ith_tree_scores = lfi_scores[:, :, i] oob_indices = np.unique(self.oob_indices[i]) if evaluate_on == 'oob': local_feature_importances[oob_indices, :, i] = \ ith_tree_scores[oob_indices, :] elif evaluate_on == 'inbag': inbag_indices = np.arange(X.shape[0]) inbag_indices = np.setdiff1d(inbag_indices, oob_indices) local_feature_importances[inbag_indices, :, i] = \ ith_tree_scores[inbag_indices, :] else: local_feature_importances[:, :, i] = ith_tree_scores if ranking: local_feature_importances = np.abs(local_feature_importances) rank_matrix = np.zeros_like(local_feature_importances) for i in range(local_feature_importances.shape[-1]): lfi_treei = local_feature_importances[:,:,i] indices_of_zero_columns = np.where(np.all(lfi_treei==0, axis=0))[0] lfi_treei[:, indices_of_zero_columns] = -1 ranks = np.argsort(np.argsort(lfi_treei, kind="stable"), kind = "stable") ranks = np.array(ranks, dtype=np.float32) ranks[:, indices_of_zero_columns] = np.nan rank_matrix[:,:,i] = ranks local_feature_importances = rank_matrix local_feature_importances = np.nanmean(local_feature_importances, axis=-1) local_feature_importances[np.isnan(local_feature_importances)] = 0 return local_feature_importances def _get_LFI_subtract_intercept(self, X, njobs): """ Compute per-tree LMDI+ scores for each sample in X. Parameters: ---------- X : np.ndarray Input feature matrix of shape (n_samples, n_features). njobs : int Number of parallel jobs to use for prediction (if supported). Returns: ------- LFIs : np.ndarray Matrix of shape (n_samples, n_features, n_trees) containing the per-tree local feature importance scores. """ LFIs = np.zeros((X.shape[0],X.shape[1],len(self.tree_explainers))) for i, tree_explainer in enumerate(self.tree_explainers): blocked_data_ith_tree = self.rf_plus_model.transformers_[i].transform(X) if self.rf_plus_model._task == "classification": ith_partial_preds = tree_explainer.predict_partial_subtract_intercept(blocked_data_ith_tree, njobs=njobs) else: ith_partial_preds = tree_explainer.predict_partial_subtract_intercept(blocked_data_ith_tree, njobs=njobs) ith_partial_preds = np.array([ith_partial_preds[j] for j in range(X.shape[1])]).T LFIs[:,:,i] = ith_partial_preds return LFIsMethods
def get_lmdi_plus_scores(self, X, y=None, njobs=1, ranking=False)-
Compute LMDI+ scores for each sample in X.
If
yis provided, the evaluation is assumed to be on the training set, and ifyis None, it assumes LFI computation is being performed on unseen test data.Parameters:
X : np.ndarray Input feature matrix of shape (n_samples, n_features). y : np.ndarray or None Target values for the training set or None for test data. njobs : int Number of parallel jobs to use for prediction (if supported). ranking : bool If True, converts the LFI scores to feature rankings per sample.
Returns:
local_feature_importances : np.ndarray Matrix of shape (n_samples, n_features) containing the averaged LMDI+ scores across trees.
Expand source code
def get_lmdi_plus_scores(self, X, y = None, njobs = 1, ranking = False): """ Compute LMDI+ scores for each sample in X. If `y` is provided, the evaluation is assumed to be on the training set, and if `y` is None, it assumes LFI computation is being performed on unseen test data. Parameters: ---------- X : np.ndarray Input feature matrix of shape (n_samples, n_features). y : np.ndarray or None Target values for the training set or None for test data. njobs : int Number of parallel jobs to use for prediction (if supported). ranking : bool If True, converts the LFI scores to feature rankings per sample. Returns: ------- local_feature_importances : np.ndarray Matrix of shape (n_samples, n_features) containing the averaged LMDI+ scores across trees. """ local_feature_importances = np.full((X.shape[0], X.shape[1],len(self.tree_explainers)), np.nan) if y is None: evaluate_on = None else: evaluate_on = self.evaluate_on lfi_scores = self._get_LFI_subtract_intercept(X, njobs) for i in range(lfi_scores.shape[-1]): ith_tree_scores = lfi_scores[:, :, i] oob_indices = np.unique(self.oob_indices[i]) if evaluate_on == 'oob': local_feature_importances[oob_indices, :, i] = \ ith_tree_scores[oob_indices, :] elif evaluate_on == 'inbag': inbag_indices = np.arange(X.shape[0]) inbag_indices = np.setdiff1d(inbag_indices, oob_indices) local_feature_importances[inbag_indices, :, i] = \ ith_tree_scores[inbag_indices, :] else: local_feature_importances[:, :, i] = ith_tree_scores if ranking: local_feature_importances = np.abs(local_feature_importances) rank_matrix = np.zeros_like(local_feature_importances) for i in range(local_feature_importances.shape[-1]): lfi_treei = local_feature_importances[:,:,i] indices_of_zero_columns = np.where(np.all(lfi_treei==0, axis=0))[0] lfi_treei[:, indices_of_zero_columns] = -1 ranks = np.argsort(np.argsort(lfi_treei, kind="stable"), kind = "stable") ranks = np.array(ranks, dtype=np.float32) ranks[:, indices_of_zero_columns] = np.nan rank_matrix[:,:,i] = ranks local_feature_importances = rank_matrix local_feature_importances = np.nanmean(local_feature_importances, axis=-1) local_feature_importances[np.isnan(local_feature_importances)] = 0 return local_feature_importances