Addressing Bias in Machine Learning Models Using Marginal Contribution

Abstract

Fairness in AI is a crucial aspect of modern machine learning. We focus on the correlation between sensitive and non-sensitive variables which plays a trick when learning models, known as the red-lining effect.In this paper, we present a new algorithm for handling the correlation in machine learning models using marginal contribution. We first clarify a necessarily and sufficient condition between marginal contribution and the independence of sensitive and non-sensitive variables, and then use this condition to develop an algorithm for addressing correlation in models. We then evaluate its performance through empirical experiments.