Optimization approaches for fairness-aware machine learning

In recent years, artificial intelligence (AI) and machine learning (ML) technologies have been used in high-stakes decision making systems like lending decision, employment screening, criminal justice sentencing and resource allocation. However, a new challenge arising with these AI systems is avoiding the unfairness introduced by the systems that lead to discriminatory decisions for protected groups defined by some sensitive variables (e.g., age, race, gender). Among the techniques for improving the fairness of AI systems, the optimization-based method, which trains a model through optimizing its prediction performance subject to fairness constraints, is most popular because of its intuitive idea and the Pareto efficiency it guarantees when trading off prediction performance against fairness. However, most existing optimization models are formulated using threshold-dependent metrics for performance and fairness, which is insufficient to mitigate minority bias and fails to guarantee fairness if the thresholds used in training and testing are inconsistent. Moreover, most existing approaches reformulate the fairness constraints as regularization terms, which requires tuning a regularization parameter and may produce an infeasible solution that violates the fairness constraints. Therefore, this thesis focuses on the design of new threshold-agnostic fairness metrics and the development of efficient numerical algorithms for solving the optimization problem with threshold-agnostic objective and constraints directly with theoretical guarantee.