[Paper Review] Counterfactual Fairness

Matt J. Kusner, Joshua R. Loftus+2 more|arXiv (Cornell University)|Mar 20, 2017

Ethics and Social Impacts of AI31 references868 citations

TL;DR

Introduces counterfactual fairness, a causal framework ensuring predictions are invariant to protected attributes under counterfactual changes, and presents an algorithm to learn such fair predictors, illustrated on law school success prediction.

ABSTRACT

Machine learning can impact people with legal or ethical consequences when it is used to automate decisions in areas such as insurance, lending, hiring, and predictive policing. In many of these scenarios, previous decisions have been made that are unfairly biased against certain subpopulations, for example those of a particular race, gender, or sexual orientation. Since this past data may be biased, machine learning predictors must account for this to avoid perpetuating or creating discriminatory practices. In this paper, we develop a framework for modeling fairness using tools from causal inference. Our definition of counterfactual fairness captures the intuition that a decision is fair towards an individual if it is the same in (a) the actual world and (b) a counterfactual world where the individual belonged to a different demographic group. We demonstrate our framework on a real-world problem of fair prediction of success in law school.

Motivation & Objective

• Motivate the need for fairness-aware predictions in high-stakes domains with historical biases.
• Propose a formal, causal definition of counterfactual fairness based on Pearl's framework.
• Develop an algorithm to learn predictive models that satisfy counterfactual fairness by leveraging latent variables.
• Illustrate the approach with a real-world-like case study predicting law school success.
• Discuss implications, design choices, and limitations of causal fairness in practice.

Proposed method

• Adopt a causal model (U,V,F) with protected attributes A, observed X, and outcome Y.
• Define counterfactual fairness: hatY is fair if its distribution is invariant under A changing counterfactually (A ← a′) given X and A.
• Propose a learning algorithm FairLearning that samples latent U from P_M(U | X, A) via MCMC and trains g_theta(U, X_nsucc(A)) to minimize predictive loss.
• Provide a prediction-time estimate as E[hatY(U*, X_nsucc(A*)) | X*, A*].
• Show connections to existing fairness notions and argue for explicit causal modeling over purely statistical criteria.

Experimental results

Research questions

• RQ1 How to define a fairness-consistent predictor in a causal setting where protected attributes causally influence observed data?
• RQ2 Can we construct predictors whose outcomes are invariant to counterfactual changes in protected attributes while maintaining predictive usefulness?
• RQ3 What are practical learning strategies to approximate counterfactuals given a causal model and finite data?
• RQ4 How does counterfactual fairness relate to and potentially reconcile other fairness criteria like Equal Opportunity and Predictive Parity?

Key findings

• Counterfactual fairness provides an individual-level fairness notion that remains invariant to counterfactual changes in protected attributes along causal paths.
• An algorithm (FairLearning) is proposed to learn fair predictors by grounding them in latent variables U and non-descendants of A, using MCMC to approximate counterfactual expectations.
• In a law school success illustration, counterfactually fair models trade some predictive accuracy for fairness, with Level 1–3 approaches offering progressively stronger assumptions and different RMSE outcomes.
• Using latent fair variables (Level 2) or additive-error models (Level 3) yields different predictive performance versus Full and Unaware baselines, demonstrating the fairness-accuracy trade-off.
• Counterfactual fairness helps address fairness paradoxes where traditional criteria may conflict when Y and A are associated.