Date of Graduation

5-2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Engineering (PhD)

Degree Level

Graduate

Department

Computer Science & Computer Engineering

Advisor/Mentor

Wu, Xintao

Committee Member

Gauch, Susan E.

Second Committee Member

Liu, Xiao

Third Committee Member

Zhang, Lu

Keywords

Artificial Intelligence; Bandit; Causal Inference; Machine Learning; Recommendation

Abstract

Recommender systems provide personalized services for users seeking information and play an increasingly important role in online applications. While most research papers focus on inventing machine learning algorithms to fit user behavior data and maximizing predictive performance in recommendation, it is also very important to develop fairness-aware machine learning algorithms such that the decisions made by them are not only accurate but also meet desired fairness requirements. In personalized recommendation, although there are many works focusing on fairness and discrimination, how to achieve user-side fairness in bandit recommendation from a causal perspective still remains a challenging task. Besides, the deployed systems utilize user-item interaction data to train models and then generate new data by online recommendation. This feedback loop in recommendation often results in various biases in observational data. The goal of this dissertation is to address challenging issues in achieving causal fairness in recommender systems: achieving user-side fairness and counterfactual fairness in bandit-based recommendation, mitigating confounding and sample selection bias simultaneously in recommendation and robustly improving bandit learning process with biased offline data. In this dissertation, we developed the following algorithms and frameworks for research problems related to causal fairness in recommendation. • We developed a contextual bandit algorithm to achieve group level user-side fairness and two UCB-based causal bandit algorithms to achieve counterfactual individual fairness for personalized recommendation; • We derived sufficient and necessary graphical conditions for identifying and estimating three causal quantities under the presence of confounding and sample selection biases and proposed a framework for leveraging the causal bound derived from the confounded and selection biased offline data to robustly improve online bandit learning process; • We developed a framework for discrimination analysis with the benefit of multiple causes of the outcome variable to deal with hidden confounding; • We proposed a new causal-based fairness notion and developed algorithms for determining whether an individual or a group of individuals is discriminated in terms of equality of effort.

Share

COinS