def audit_tree(d, features, outcome, protected):
kernel = DotProduct() + WhiteKernel()
# create learners
for varname, attribute_list in protected.items():
# remove protected from features
features_audit = [feat for feat in features if feat != varname]
learner_list = []
# split (70/30)
train = d.loc[np.random.choice(d.index,
int(0.7 * len(d)),
replace=False)]
test = d.drop(train.index)
for attribute in attribute_list:
data = train[train[varname] == attribute]
learner = GaussianProcessRegressor(kernel=kernel, random_state=0)
#learner = DecisionTreeClassifier()
learner.name = attribute
learner = fit(data, features_audit, outcome, learner)
learner_list.append(pickle.dumps(learner))
score = np.inf
for attribute in attribute_list:
data = test[test[varname] == attribute]
results = predict(learner_list, data, features_audit, outcome)
#results_confusion = results[["predict_%s"%var for var in attribute_list]].mean(axis=1)
results_confusion = np.array(
results[["predict_%s" % var for var in attribute_list]])
results_variance = np.ones(len(results_confusion))
results_confusion[results_confusion < 0.5] = 0
results_confusion[results_confusion >= 0.5] = 1
print(
accuracy_score(np.array(data[outcome]), results_confusion[:,
0]))
print(
accuracy_score(np.array(data[outcome]), results_confusion[:,
1]))
results_difference = np.absolute(results_confusion[:, 0] -
results_confusion[:, 1])
for var in attribute_list:
results_variance = results_variance * np.array(
results["confidence_%s" % var])
if (((results_difference / results_variance).sum()) *
results_variance.sum()) < score:
score = ((results_difference /
results_variance).sum()) * results_variance.sum()
return score
