INTERPRETABILITY AND FAIRNESS IN MACHINE LEARNING: A FORMAL METHODS APPROACH

Abstract

The significant success of machine learning in past decades has led to a host of applications of algorithmic decision-making in different safety-critical domains. The high-stake predictions of machine learning in medical, law, education, transportation and so on have far-reaching consequences on the end-users. Consequently, there has been a call for the regulation of machine learning by defining and improving the interpretability, fairness, robustness, and privacy of predictions. In this thesis, we focus on the interpretability and fairness aspects of machine learning, particularly on learning interpretable rule-based classifiers, verifying fairness, and interpreting sources of unfairness. Prior studies aimed for these problems are limited by either scalability or accuracy or both. To alleviate these limitations, we integrate formal methods and automated reasoning with interpretability and fairness in machine learning and provide scalable and accurate solutions to the underlying problems.