import sklearn.preprocessing
import numpy as np
from aif360.algorithms import Transformer
class LimeEncoder(Transformer):
"""Tranformer for converting aif360 dataset to LIME dataset and vice versa.
(LIME - Local Interpretable Model-Agnostic Explanations) [2]_
See for details/usage:
https://github.com/marcotcr/lime
References:
.. [2] M.T. Ribeiro, S. Singh, and C. Guestrin, '"Why should I trust
you?" Explaining the predictions of any classifier.'
https://arxiv.org/pdf/1602.04938v1.pdf
"""
def __init__(self):
super(LimeEncoder, self).__init__()
def fit(self, dataset):
"""Take an aif360 dataset and save all relevant metadata as well as
mappings needed to transform/inverse_transform the data between aif360
and lime.
Args:
dataset (BinaryLabelDataset): aif360 dataset
Returns:
LimeEncoder: Returns self.
"""
self.s_feature_names_with_one_hot_encoding = dataset.feature_names
df, df_dict = dataset.convert_to_dataframe(de_dummy_code=True)
dfc = df.drop(dataset.label_names[0], axis=1) # remove label (class) column
self.s_feature_names = list(dfc.columns) # create list of feature names
self.s_data = dfc.values # create array of feature values
# since categorical features are 1-hot-encoded and their names changed,
# the set diff gives us the list of categorical features as non-
# categorical feature names are not changed
self.s_categorical_features = list(set(self.s_feature_names)
- set(self.s_feature_names_with_one_hot_encoding))
self.s_protected_attribute_names = dataset.protected_attribute_names
# add protected attribute names to the list of categorical features
self.s_categorical_features = self.s_categorical_features \
+ self.s_protected_attribute_names
self.s_labels = df[dataset.label_names[0]] # create labels
# following 3 lines are not really needed
# using to create s_class_names..can do so manually as well ...array([ 0., 1.])
s_le = sklearn.preprocessing.LabelEncoder()
s_le.fit(self.s_labels)
# self.s_labels = s_le.transform(self.s_labels)
self.s_class_names = s_le.classes_
# convert s_categorical_features to a list of array indexes in
# s_feature_names corresponding to categorical features
# (NOTE - does not included protected attributes)
self.s_categorical_features = [self.s_feature_names.index(x)
for x in self.s_categorical_features]
# map all the categorical features to numerical values and store the
# mappings in s_categorical_names
self.s_categorical_names = {}
for feature in self.s_categorical_features:
self.le = sklearn.preprocessing.LabelEncoder()
self.le.fit(self.s_data[:, feature])
#self.s_data[:, feature] = le.transform(self.s_data[:, feature])
self.s_categorical_names[feature] = self.le.classes_
return self
def transform(self, aif360data):
"""Take aif360 data array and return data array that is lime encoded
(numeric array in which categorical features are NOT one-hot-encoded).
Args:
aif360data (np.ndarray): Dataset features
Returns:
np.ndarray: LIME dataset features
"""
tgtNumRows = aif360data.shape[0]
tgtNumcolumns = len(self.s_feature_names)
limedata = np.zeros(shape=(tgtNumRows, tgtNumcolumns))
# non_categorical_features = list(set(self.s_feature_names) & set(self.s_feature_names_with_one_hot_encoding))
for rw in range(limedata.shape[0]):
for ind, feature in enumerate(self.s_feature_names):
if ind in self.s_categorical_features:
# tranform the value since categorical feature except if it
# is also a protected attribute
if feature in self.s_protected_attribute_names:
# just copy the value as is
limedata[rw, ind] = aif360data[rw, self.s_feature_names_with_one_hot_encoding.index(feature)]
else:
possible_feature_values = self.s_categorical_names[ind]
for indc in range(len(possible_feature_values)):
cval = possible_feature_values[indc]
colName = feature + "=" + cval
if (aif360data[rw][self.s_feature_names_with_one_hot_encoding.index(colName)] == 1.0):
limedata[rw][ind] = indc
else:
# just copy the value as is
limedata[rw, ind] = aif360data[rw, self.s_feature_names_with_one_hot_encoding.index(feature)]
return limedata
def inverse_transform(self, limedata):
"""Take data array that is lime encoded (that is, lime-compatible data
created by this class from a given aif360 dataset) and return data array
consistent with the original aif360 dataset.
Args:
limedata (np.ndarray): Dataset features
Returns:
np.ndarray: aif360 dataset features
"""
tgtNumRows = limedata.shape[0]
tgtNumcolumns = len(self.s_feature_names_with_one_hot_encoding)
aif360data = np.zeros(shape=(tgtNumRows, tgtNumcolumns))
for rw in range(aif360data.shape[0]):
for ind, feature in enumerate(self.s_feature_names):
# s_categorical_features has list of indexes into
# s_feature_names for categorical features
if ind in self.s_categorical_features:
if feature in self.s_protected_attribute_names:
# just copy the value as is
aif360data[rw, self.s_feature_names_with_one_hot_encoding.index(feature)] = limedata[rw, ind]
else:
# s_categorical_names[ind] has mapping of categorical to
# numerical values i.e. limedata[rw, ind] is index of
# this array. value is string val
new_feature = feature + '=' + self.s_categorical_names[ind][int(limedata[rw, ind])]
# categorical feature:
aif360data[rw, self.s_feature_names_with_one_hot_encoding.index(new_feature)] = 1.0
else: # just copy value
aif360data[rw, self.s_feature_names_with_one_hot_encoding.index(feature)] = limedata[rw, ind]
return aif360data