Operationalizing-Individual-Fairness

Implementation of "Operationalizing Individual Fairness with Pairwise Fair Representations"
Link to Google Slides presentation

Setup

Dataset	Number of Records (Paper)	Number of Records (Our Implementation)	Number of Features (Paper)	Number of Features ( Our Implementation )	True Rank	Protected Attribute
Compas	8803	6903	429	456	117	Race_African-American

Requirements

A list of packages required to run is mentioned in requirements.txt file.

Results

Reproduced Results :

Original Representation	Pairwise Fair Representation (Gamma = 0.5)
Accuracy: 69.56%	Accuracy: 66.01%
ROC-AUC score: 69.04%	ROC-AUC score: 65.68
Positive Prediction Rate for African Americans: 0.419	Positive Prediction Rate for African Americans: 0.434
Positive Prediction Rate for Non-African Americans: 0.406	Positive Prediction Rate for Non-African Americans: 0.5
Prediction Error Rate for African Americans: 0.323	Prediction Error Rate for African Americans: 0.34
Prediction Error Rate for Non-African Americans: 0.304	Positive Prediction Rate for Non-African Americans: 0.75
False Positive Rate for African Americans: 0.3	False Positive Rate for African Americans: 0.285
False Positive Rate for Non-African Americans: 0.226	False Positive Rate for Non-African Americans: 0.667

NOTE

Concerns Faced in the experimentation :

1. We did not experiment on the Crime and Communities due to the requirement of of ratings columns for the neighbourhoods from niche.com which is not available as a public dataset.
2. We performed trial and error methods to tune parameters k( for kth quantile ) and p( for p nearest neighbours ) for maximum value of AUC score.

Team Members

Name	Roll Number
Harshvardhan Srivastava	17EE10058
Sanket Kumar Singh	17EE30016