def randomOneSideMassageData(examples, protectedIndex, protectedValue):
bias = signedStatisticalParity(examples, protectedIndex, protectedValue)
print("Initial bias:", bias)
favored_trait = zeroOneSign(bias)
#break up data by label and by the value of the protected trait
favored_data = [(x,label) for x,label in examples if x[protectedIndex]==favored_trait]
nonfavored_data = [(x,label) for x,label in examples if x[protectedIndex]!=favored_trait]
favored_data_positive = [pt for pt in favored_data if pt[1]==1]
nonfavored_data_negative = [pt for pt in nonfavored_data if pt[1]==-1]
#calculate number of labels to flip from -1 to +1 on the nonfavored side
num_nonfavored_positive = len(nonfavored_data)-len(nonfavored_data_negative)
num_to_flip = math.floor((len(nonfavored_data)*len(favored_data_positive) - len(favored_data)*num_nonfavored_positive)/len(favored_data))
print("Number of labels flipped:",num_to_flip)
to_flip_to_pos = sample(nonfavored_data_negative, num_to_flip)
flipped_examples = []
for data in examples:
if data in to_flip_to_pos:
flipped_examples.append((data[0],-1*data[1]))
else:
flipped_examples.append(data)
return flipped_examples
def thresholdRelabel(h, trainingData, protectedIndex, protectedValue,
hypotheses, weights, threshold):
bias = signedStatisticalParity(trainingData, protectedIndex, protectedValue, h)
biasedClass = 1 - zeroOneSign(bias)
def relabel(pt):
proposedLabel = h(pt)
if (pt[protectedIndex] == biasedClass and
absMargin(pt, hypotheses, weights) < threshold):
return -proposedLabel
else:
return proposedLabel
return relabel
def thresholdRelabel(h, trainingData, protectedIndex, protectedValue,
hypotheses, weights, threshold):
bias = signedStatisticalParity(trainingData, protectedIndex,
protectedValue, h)
biasedClass = 1 - zeroOneSign(bias)
def relabel(pt):
proposedLabel = h(pt)
if (pt[protectedIndex] == biasedClass
and absMargin(pt, hypotheses, weights) < threshold):
return -proposedLabel
else:
return proposedLabel
return relabel
def randomOneSideRelabelData(h, trainingData, protectedIndex, protectedValue):
bias = signedStatisticalParity(trainingData, protectedIndex, protectedValue, h)
favored_trait = zeroOneSign(bias)
nonfavored_data = [(feats,label) for feats,label in trainingData if not feats[protectedIndex]==favored_trait]
NF, NFn = len(nonfavored_data), len([1 for x,label in nonfavored_data if h(x)==-1])
p = NF*abs(bias)/NFn
def relabeledClassifier(point):
origClass = h(point)
if point[protectedIndex] != favored_trait and origClass == -1:
if random() < p:
return -origClass
else:
return origClass
else:
return origClass
return relabeledClassifier
def randomOneSideRelabelData(h, trainingData, protectedIndex, protectedValue):
bias = signedStatisticalParity(trainingData, protectedIndex,
protectedValue, h)
favored_trait = zeroOneSign(bias)
nonfavored_data = [(feats, label) for feats, label in trainingData
if not feats[protectedIndex] == favored_trait]
NF, NFn = (len(nonfavored_data),
len([1 for x, label in nonfavored_data if h(x) == -1]))
p = NF * abs(bias) / NFn
def relabeledClassifier(point):
origClass = h(point)
if point[protectedIndex] != favored_trait and origClass == -1:
if random() < p:
return -origClass
else:
return origClass
else:
return origClass
return relabeledClassifier