def generate_neighborhood_data(self,
sample,
predict_fn,
distance_metric='euclidean',
n_samples=500,
seed=1,
**kwargs):
'''Generate neighborhood data for a given point (currently using LIME)
Args:
train_data: Training data predict_fn was trained on
sample: Observed sample
predict_fn: Black box predictor to predict all points
distance_metric: Distance metric used for weights
n_samples: Number of samples to generate
Returns:
neighor_data (xs around sample),
weights (weights of instances in xs),
neighor_data_labels (ys around sample, corresponding to xs)
'''
from lime.lime_tabular import LimeTabularExplainer
e = LimeTabularExplainer(
self.train_data,
categorical_features=self.categorical_features,
discretize_continuous=False)
_, neighbor_data = e._LimeTabularExplainer__data_inverse(
sample, n_samples)
scaled_data = (neighbor_data - e.scaler.mean_) / e.scaler.scale_
return (*self._data(neighbor_data, scaled_data, distance_metric,
predict_fn), sample)
def generate_neighborhood_data(self,
sample,
predict_fn,
distance_metric='euclidean',
n_samples=500,
seed=1,
**kwargs):
"""Generate neighborhood data for a given point (currently samples training data)
Args:
sample: Observed sample
predict_fn: Black box predictor to predict all points
distance_metric: Distance metric used for weights
n_samples: Number of samples to generate
Returns:
neighor_data (xs around sample),
weights (weights of instances in xs),
neighor_data_labels (ys around sample, corresponding to xs)
"""
from lime.lime_tabular import LimeTabularExplainer
categorical_features = None
if self.categorical_features is not None:
cfi = itertools.chain.from_iterable
categorical_features = list(
cfi(self.feature_map[c] for c in self.categorical_features))
e = LimeTabularExplainer(self.train_data,
categorical_features=categorical_features,
discretize_continuous=False,
random_state=check_random_state(seed))
_, neighbor_data = e._LimeTabularExplainer__data_inverse(
sample, n_samples)
scaled_data = (neighbor_data - e.scaler.mean_) / e.scaler.scale_
predict_data = self._apply_decode(neighbor_data)
return (*self._data(neighbor_data, scaled_data, predict_data,
distance_metric, predict_fn), sample)
