ImpartialGAN: Fair and Unbiased Classification

Abstract

In this age of big data, one of the key concerns in the recent days has been bias present in the data and hence the need to ensure data fairness. There is a need to ensure that bias in the data does not reflect in the models decision which in turn treats people from certain race, gender, sexual or political orientation unfairly and differently. The goal of fair data generation is to remove any prejudice which might be present in the data towards any specific demographic group. This is particularly of interest in decision making scenarios like financial lending, hiring, pretrial and immigration detention, health care, social services, and education where the system might favor one race and is biased towards the other. In this paper, we propose ImpartialGAN to generate fair synthetic data from real data. The generated data is not only fair and free from bias but also ensures a good data utility while preserving data privacy. Hence this generated data can be used in place of real data for predictive analytics. In our experiments on UCI Adult dataset, we achieved 83.43 % accuracy on real data while keeping the risk difference for synthetic data at 0.0063, which indicates that our classifier is fair and unbiased.