Fairness and Algorithmic Decision Making

This course examines the greater context under which the practice of Data Science exists and explores concrete ways these issues surface in technical work. Two parallel threads run through the quarter:

  • Readings introduce students to different frameworks for understanding how individuals relate to social institutions.
  • Identifying how these concepts arise in the “life of a Data Scientist” and using them to propose and critique potential solutions.

Course Outcomes:

  • Differentiate between descriptive, normative, and applied ethics; identify arguments in case-studies that rely on each.
  • Attain a basic understanding of fundamental concepts like fairness, justice, and equality.
  • Articulate how notions of fairness are quantified and understand their relative advantages and disadvantages (parity measures and their impossibility results; counterfactuals and causality).
  • Study issues of fairness in case studies, identifying possible causes within a social/legal context and consider possible (quantitative) actions using an ethical framework.
  • Consider the impacts of the amplification of bias at an institutional level, possible social actions, and the limits of quantification.
  • Identify unfairness in data using quantitative techniques with code and identify the limitations of these techniques in case-studies.

Course Format:

  • Lecture format
  • Readings + (iClicker) Quizzes
  • Two papers (involving code/statistical analysis).
  • Final Exam

Ungraded coding homework will help students complete their papers and prepare for the final.

Course Topics

Much of the course draws on the work collated in the book Fairness and Machine Learning, as well as pioneering (graduate) courses at Berkeley, Cornell, and Princeton.

Fundamental Concepts

  • Different ethical frameworks (descriptive, normative, applied).
  • Definitions: Fairness, justice, equality
  • Institutional contexts under which these issues appear (e.g. legal, social)
  • How algorithms engage in “decision making” (automated actions, statistical analyses, delivery of information).

Notions of fairness

  • Notions of fairness/justice (individual vs group).
  • Observational notions of fairness (parity measures).
  • Impossibility results
  • Identifying unfairness in data with code.

How unfair behaviors manifest in models

  • “Blame it on the training data” and “confusing proxies for groundtruth”.
  • Aspirational fairness (how should a model behave?).
  • Feedback loops: amplification of bias.
  • Identifying potential causes of unfairness in real-world examples.
  • Goodhart’s Law.

Counterfactual notions of fairness (causality)

  • Quantifying concepts from Rawls’s “Theory of Justice”
  • Identifying cause using natural experiments.
  • Counterfactual fairness and machine learning models.

Limits of Quantification

  • “Which groups should be considered vulnerable, and can they be ‘discovered’ algorithmically?”
  • Does quantifying fairness require them to be independent of social context? (Third Party APIs).
  • Transparency vs. Personalization.