Testing, Debugging, and Repairing Individual Discrimination in Machine Learning Models

Abstract

Trustworthy AI is crucial for the broad adoption of AI systems. An important question is, therefore, how to ensure this trustworthiness. The absence of algorithmic bias is a crucial attribute for an AI system to be considered trustworthy. In this book chapter, we address various problems related to the detection and mitigation of algorithmic bias in machine learning models, specifically individual discrimination. A model shows individual discrimination if two instances, predominantly differing in protected attributes like race, gender, or age, produce different decision outcomes. In a black-box setting, detecting individual discrimination requires extensive testing. We present a methodology that enables the automatic generation of test inputs with a high likelihood of identifying individual discrimination. Our approach unites the power of two widely recognized techniques, symbolic execution and local explainability, to generate test cases effectively. We further address the problem of localizing individual discrimination failures required for effective model debugging. We introduce a notion called Region of Individual Discrimination or RID, which is an interpretable region in the feature space, described by simple predicates on features, aiming to contain all the discriminatory instances. This region essentially captures the positive correlation between discriminatory instances and features. Finally, we describe a model repair algorithm that aims to repair individual discrimination in the model by effectively creating retraining data based on RID. We empirically show that our approaches are effective for assuring individual fairness of machine learning models.