Empirical analysis of Information Theory point of view on Equal Odds Fairness Measure

Abstract

Algorithmic decision-making is facing a rising concern of unwanted discrimination. Biases that exist in decision-making, data collection, or in the attributes can lead to unfair decisions, and consequently to legal issues. In this paper, we investigate the information theory definition of equal odds and its decomposition to independence, balance, and accuracy on equal odds. Based on those definitions we design three linear regularization terms and observe how these regularizations influence each other. The results on three datasets, namely Adult, COMPAS, and German, suggest that each regularization improves fairness in terms of disparate impact, balance, and equal odds. However, separation regularization performs the best in terms of equal odds. An interesting finding is that the proposed regularizations do not reduce accuracy, which is in accordance to the information theory point of view on equal odds.

Sandro Radovanović

Assistant Professor at University of Belgrade

My research interests include machine learning, development and design of decision support systems, decision theory, and fairness and justice concepts in algorithmic decision making.