Added algorithms and test file for Celis et al. paper. (Trusted-AI#21)

* Added algorithms and test file for Celis et al. paper.

* Expanded example and changed stats function

* Made small changes to main file and example and added unit test

* Fixed bugs in notebook and added unbiased classifier for comparison

* Modified test file and made small clean-up changes to main files. Changed README to include link to Celis et al paper.

* Small change to check build

README.md

@@ -24,6 +24,7 @@ Get in touch with us on [Slack](https://aif360.slack.com) (invitation [here](htt
 * Calibrated Equalized Odds Postprocessing ([Pleiss et al., 2017](https://papers.nips.cc/paper/7151-on-fairness-and-calibration))
 * Learning Fair Representations ([Zemel et al., 2013](http://proceedings.mlr.press/v28/zemel13.html))
 * Adversarial Debiasing ([Zhang et al., 2018](http://www.aies-conference.com/wp-content/papers/main/AIES_2018_paper_162.pdf))
+* Meta-Algorithm for Fair Classification ([Celis et al.. 2018](https://arxiv.org/abs/1806.06055))
 
 ## Supported fairness metrics
 

aif360/algorithms/inprocessing/__init__.py

@@ -1,3 +1,4 @@
 from aif360.algorithms.inprocessing.adversarial_debiasing import AdversarialDebiasing
 from aif360.algorithms.inprocessing.art_classifier import ARTClassifier
 from aif360.algorithms.inprocessing.prejudice_remover import PrejudiceRemover
+from aif360.algorithms.inprocessing.meta_fair_classifier import MetaFairClassifier

aif360/algorithms/inprocessing/celisMeta/FalseDiscovery.py

@@ -0,0 +1,150 @@
+import os,sys
+from scipy.stats import multivariate_normal
+import scipy.stats as st
+import numpy as np
+import math
+
+import site
+site.addsitedir('.')
+
+from .General import *
+from . import utils as ut
+
+
+class FalseDiscovery(General):
+
+	def getExpectedGrad(self, dist_params, params, samples, mu,  z_0, z_1, a, b):
+		u_1, u_2, l_1, l_2 = params[0], params[1], params[2], params[3]
+		a, b = a[0], b[0]
+		res1 = []
+		res2 = []
+		res3 = []
+		res4 = []
+		for x in samples:
+				temp = np.append(np.append(x, 1), 1)
+				prob_1_1 = ut.getProbability(dist_params, temp)
+
+				temp = np.append(np.append(x, -1), 1)
+				prob_m1_1 = ut.getProbability(dist_params, temp)
+
+				temp = np.append(np.append(x, 1), 0)
+				prob_1_0 = ut.getProbability(dist_params, temp)
+
+				temp = np.append(np.append(x, -1), 0)
+				prob_m1_0 = ut.getProbability(dist_params, temp)
+
+
+				prob_y_1 = (prob_1_1 + prob_1_0) / (prob_1_1 + prob_1_0 + prob_m1_0 + prob_m1_1)
+				#print(prob_y_1)
+
+				prob_z_0 = (prob_m1_0 + prob_1_0) / (prob_1_1 + prob_1_0 + prob_m1_0 + prob_m1_1)
+				prob_z_1 = (prob_m1_1 + prob_1_1) / (prob_1_1 + prob_1_0 + prob_m1_0 + prob_m1_1)
+
+				probc_m1_0 = prob_m1_0 / (prob_1_1 + prob_1_0 + prob_m1_0 + prob_m1_1)
+				probc_m1_1 = prob_m1_1 / (prob_1_1 + prob_1_0 + prob_m1_0 + prob_m1_1)
+
+				c_0 = prob_y_1 - 0.5
+				c_1 = u_1 * (probc_m1_0 - a*prob_z_0) + u_2 * (probc_m1_1 - a*prob_z_1)
+				c_2 = l_1 * (- probc_m1_0 + b*prob_z_0) + l_2 * (- probc_m1_1 + b*prob_z_1)
+
+				t = math.sqrt((c_0 + c_1 + c_2)*(c_0 + c_1 + c_2) + mu*mu)
+				t1 = (c_0 + c_1 + c_2) * (probc_m1_0 - a*prob_z_0)/t
+				t2 = (c_0 + c_1 + c_2) * (probc_m1_1 - a*prob_z_1)/t
+				t3 = (c_0 + c_1 + c_2) * (- probc_m1_0 + b*prob_z_0)/t
+				t4 = (c_0 + c_1 + c_2) * (- probc_m1_1 + b*prob_z_1)/t
+				#print(t1,t2)
+				res1.append(t1)
+				res2.append(t2)
+				res3.append(t3)
+				res4.append(t4)
+
+		return [np.mean(res1), np.mean(res2), np.mean(res3), np.mean(res4)]
+
+	def getValueForX(self, dist_params, a,b, params, samples,  z_0, z_1, x, flag):
+				u_1, u_2, l_1, l_2 = params[0], params[1], params[2], params[3]
+				#print (params)
+				a, b = a[0], b[0]
+
+				temp = np.append(np.append(x, 1), 1)
+				prob_1_1 = ut.getProbability(dist_params, temp)
+
+				temp = np.append(np.append(x, -1), 1)
+				prob_m1_1 = ut.getProbability(dist_params, temp)
+
+				temp = np.append(np.append(x, 1), 0)
+				prob_1_0 = ut.getProbability(dist_params, temp)
+
+				temp = np.append(np.append(x, -1), 0)
+				prob_m1_0 = ut.getProbability(dist_params, temp)
+
+				if (prob_1_1 + prob_1_0 + prob_m1_0 + prob_m1_1) == 0:
+					print("Probability is 0.\n")
+					return 0
+
+				prob_y_1 = (prob_1_1 + prob_1_0) / (prob_1_1 + prob_1_0 + prob_m1_0 + prob_m1_1)
+				#print(prob_y_1)
+
+				prob_z_0 = (prob_m1_0 + prob_1_0) / (prob_1_1 + prob_1_0 + prob_m1_0 + prob_m1_1)
+				prob_z_1 = (prob_m1_1 + prob_1_1) / (prob_1_1 + prob_1_0 + prob_m1_0 + prob_m1_1)
+
+
+				probc_m1_0 = prob_m1_0 / (prob_1_1 + prob_1_0 + prob_m1_0 + prob_m1_1)
+				probc_m1_1 = prob_m1_1 / (prob_1_1 + prob_1_0 + prob_m1_0 + prob_m1_1)
+
+				c_0 = prob_y_1 - 0.5
+				c_1 = u_1 * (probc_m1_0 - a*prob_z_0) + u_2 * (probc_m1_1 - a*prob_z_1)
+				c_2 = l_1 * (- probc_m1_0 + b*prob_z_0) + l_2 * (- probc_m1_1 + b*prob_z_1)
+				if flag==1:
+					print(c_0, c_1, c_2, prob_1_1 + prob_1_0 + prob_m1_0 + prob_m1_1)
+
+				# c_1 = prob_z_0/z_0
+				# c_2 = prob_z_1/z_1
+
+				t = c_0 + c_1 + c_2
+				return t
+
+	def getFuncValue(self, dist_params, a,b, params, samples,  z_0, z_1):
+		res = []
+		for x in samples:
+				t = abs(self.getValueForX(dist_params, a,b, params, samples,  z_0, z_1, x, 0))
+				res.append(t)
+
+		exp = np.mean(res)
+		return exp
+
+	def getNumOfParams(self):
+		return 4
+
+	def getGamma(self, y_test, y_res, x_control_test):
+			pos_0 = 0
+			pos_1 = 0
+
+			z1_0 = 0
+			z1_1 = 0
+			for j in range(0,len(y_test)):
+				result = y_res[j]
+
+				if result == 1 and x_control_test[j] == 0:
+					z1_0 += 1
+				if result == 1 and x_control_test[j] == 1:
+					z1_1 += 1
+
+				actual = y_test[j]
+				if result == 1 and actual == -1 and x_control_test[j] == 0:
+					pos_0 += 1
+				if result == 1 and actual == -1 and x_control_test[j] == 1:
+					pos_1 += 1
+
+			pos_0 = float(pos_0)/z1_0
+			pos_1 = float(pos_1)/z1_1
+			if pos_0 == 0 or pos_1 == 0:
+				return 0
+			else:			
+				return min(pos_0/pos_1 , pos_1/pos_0)
+
+
+if __name__ == '__main__':
+	obj = FalseDiscovery()
+	obj.testPreprocessedData()
+	#obj.testSyntheticData()
+