Learning Beyond Utility: Fairness, Explainability, and Diversity

Abstract

Machine learning models have achieved great success in real-world applications but may introduce unfairness when optimized solely for utility performance (e.g., accuracy). This dissertation explores beyond-utility aspects, focusing on fairness and its intersections with explainability and diversity. First, it identifies and mitigates fairness issues in recommendations, addressing biases in online dating, interest diversity, and dataset imbalances. Second, it introduces explanation fairness to ensure fairness in both outcomes and decision-making, proposing a framework that balances utility, traditional fairness, and explanation fairness. Third, it examines the connections between fairness and diversity, extending diversity considerations to users and analyzing interactions on both the user and item sides. Overall, this dissertation advances responsible ML practices, promoting models that are both powerful and equitable.