Algorithmic Fairness in Sequential Decision Making

Machine learning algorithms have been used on a wide range of applications, and there are growing concerns about potential biases of those algorithms. While many solutions have been proposed for addressing biases in predictions from an algorithm, there is still a gap in translating predictions to a justified decision. Moreover, even a justified and fair decision could lead to undesirable consequences with decisions create a feedback effect. While numerous solutions have been proposed for achieving fairness in one-shot decision making, there is a gap in investigating the long-term effects of sequential algorithmic decisions. In this thesis, we focus on studying algorithmic fairness in sequential decision making setting where the data comes on the fly. We first study if it is possible to translate model predictions to fair decisions. In particular, given predictions from black-box models (machine learning models or human experts), we propose an algorithm based on the classical learning-from-experts scheme to combine predictions and generate a fair and accurate decision. Our theoretical results show that approximate equalized odds can be achieved without sacrificing much regret. We also demonstrate the performance of the algorithm on real data sets commonly used by the fairness community. In the second part of the thesis, we study if enforcing static fair decisions in the sequential setting could lead to long-term equality and improvement of disadvantaged groups under feedback loop. In particular, we study the effects of repeatedly enforcing different fairness constraints at each decision time on shaping the underlying population under causal Markov Decision Models. Our results show that fairness constraints could lead to "backfire effects" which further entrenches the disparity between population groups.

Committee:

Kalyan Veeramachaneni (supervisor)

Alberto Abadie (chair)

Caroline Uhler

Alfredo Cuesta-Infante (Universidad Rey Juan Carlos)