def _weights(label_id):
scores = feature_weights[:, label_id]
_x = x
if flt_indices is not None:
scores = scores[flt_indices]
_x = mask(_x, flt_indices)
return get_top_features(feature_names, scores, top, _x)
def get_score_feature_weights(_label_id):
_score, _feature_weights = scores_weights[_label_id]
_x = x
if flt_indices is not None:
_x = mask(_x, flt_indices)
_feature_weights = mask(_feature_weights, flt_indices)
return _score, get_top_features(feature_names, _feature_weights, top,
_x)
def _weights(label_id):
coef = get_coef(clf, label_id)
_x = x
scores = _multiply(_x, coef)
if flt_indices is not None:
scores = scores[flt_indices]
_x = mask(_x, flt_indices)
return get_top_features(feature_names, scores, top, _x)
def get_score_feature_weights(_label_id):
_weights = _target_feature_weights(
weight_dicts[_label_id],
num_features=len(feature_names),
bias_idx=feature_names.bias_idx,
)
_score = _get_score(weight_dicts[_label_id])
_x = x
if flt_indices is not None:
_x = mask(_x, flt_indices)
_weights = mask(_weights, flt_indices)
return _score, get_top_features(flt_feature_names, _weights, top, _x)
def explain_predictions(self, docs, top=30):
if not isinstance(self.clf, XGBClassifier):
raise NotImplementedError
booster = self.clf.booster()
xgb_feature_names = {f: i for i, f in enumerate(booster.feature_names)}
feature_names = get_feature_names(self.clf,
self.vec,
num_features=len(xgb_feature_names))
feature_names.bias_name = '<BIAS>'
X = self.vec.transform(docs)
X = X.tocsc()
dmatrix = DMatrix(X, missing=self.clf.missing)
leaf_ids = booster.predict(dmatrix, pred_leaf=True)
tree_dumps = booster.get_dump(with_stats=True)
docs_weights = []
for i, _leaf_ids in enumerate(leaf_ids):
all_weights = _target_feature_weights(
_leaf_ids,
tree_dumps,
feature_names=feature_names,
xgb_feature_names=xgb_feature_names)[1]
weights = np.zeros_like(all_weights)
idx = X[i].nonzero()[1]
bias_idx = feature_names.bias_idx
weights[idx] = all_weights[idx]
weights[bias_idx] = all_weights[bias_idx]
docs_weights.append(weights)
weights = np.mean(docs_weights, axis=0)
feature_weights = get_top_features(feature_names=np.array(
[_prettify_feature(f) for f in feature_names]),
coef=weights,
top=top)
return Explanation(
estimator=type(self.clf).__name__,
targets=[TargetExplanation('y', feature_weights=feature_weights)],
)
def _features(label_id):
return get_top_features(state_feature_names, state_coef[label_id], top)
def _features(target_id):
coef = get_coef(reg, target_id, scale=coef_scale)
if flt_indices is not None:
coef = coef[flt_indices]
return get_top_features(feature_names, coef, top)
def _features(label_id):
coef = get_coef(clf, label_id, scale=coef_scale)
if flt_indices is not None:
coef = coef[flt_indices]
return get_top_features(feature_names, coef, top)
def _weights(label_id):
coef = get_coef(reg, label_id)
scores = _multiply(x, coef)
return get_top_features(feature_names, scores, top)
