Explainable data bias mitigation

Abstract

I propose an explainable fairness method that can not only mitigate data bias but also make humans understand its reason. Machine learning algorithms have high-risk use cases, such as hiring or lending decisions requiring fairness, accountability, and transparency. Differences in AI predictions due to sensitive attributes, such as gender, race, and age, have become an issue of fairness. While various methods to mitigate the bias of AI have been proposed, conventional methods are problematic in that they do not allow humans to intuitively understand the basis for data bias mitigation. To overcome this problem, I propose a bias mitigation method that uses explainable AI. With the proposed method, humans can understand how bias mitigation is achieved. Results of an experiment conducted by applying this method to credit scoring with the German Credit dataset show that the statistical parity difference by gender improved from -0.108 to -0.004.

The paper received Excellence Award at the 37th Annual Conference of the Japanese Society for Artificial Intelligence.