Machine learning (ML) techniques have been increasingly used to discover and reproduce patterns in human-generated data. While the hope is to improve societal outcomes with these techniques, concerns have been raised that they may inherit pre-existing biases and exhibit discrimination against certain individuals or groups. This course is about the emerging science of fairness in machine learning. It takes technical approaches to explore the sources and measures of the unfairness arising from the ML algorithms, as well as the interventions to mitigate unfairness issues. The format of the class will be a mix of lectures and research paper presentations. The course covers the following topics:

• Sources of Unfairness in ML

• Definitions of Fairness

• Tensions Between Different Fairness Notions

• Approaches to Achieving Fairness in ML

• Strategic Fairness

• Causality and Fairness

• Long-Term Impact of Short-Term Fairness Interventions

• Applications of Fairness in ML and Algorithmic Decision-Making: Lending, Criminal Justice, Labor Market, Natural Language Processing, Computer Vision, etc.

• Views of Fairness From Other Communities

• Other Ethical Issues in ML (e.g., privacy, robustness) and Their Relations with Fairness

Prerequisites: CSE 3521 or graduate standing. Students are expected to have backgrounds in machine learning/artificial intelligence, probability, and the knowledge of one programming language (e.g., Python).

Syllabus: Link

• (10%) Class participation:  two papers will be presented and discussed in each class. Students are strongly encouraged to participate in discussions actively.    

• (40%) Paper presentation: two students will present two papers in the class. Each presentation will take 30-40 minutes covering the followings:

  1. 1. Paper: backgrounds, problem formulation, model, method/algorithm/analysis, results, experiments, etc.

     2. Discussion: strengths/weaknesses, potential extensions & improvement, etc.

• (50%) Final project: students may work in a team or alone, and they can choose from the following three options:

  1. 1. Implementing a research paper on algorithmic fairness.

     2. Extending the theoretical/empirical results of previous work. 

     3. Formulating a new research problem.