[Paper Review] Avoiding Discrimination through Causal Reasoning

Niki Kilbertus, Mateo Rojas-Carulla+4 more|arXiv (Cornell University)|Jun 8, 2017

Ethics and Social Impacts of AI14 references178 citations

TL;DR

The paper reframes fairness in machine learning through causal reasoning, introducing resolving and proxy variables, and proposes intervention-based criteria and algorithms to avoid proxy and unresolved discrimination under specified causal models.

ABSTRACT

Recent work on fairness in machine learning has focused on various statistical discrimination criteria and how they trade off. Most of these criteria are observational: They depend only on the joint distribution of predictor, protected attribute, features, and outcome. While convenient to work with, observational criteria have severe inherent limitations that prevent them from resolving matters of fairness conclusively. Going beyond observational criteria, we frame the problem of discrimination based on protected attributes in the language of causal reasoning. This viewpoint shifts attention from "What is the right fairness criterion?" to "What do we want to assume about the causal data generating process?" Through the lens of causality, we make several contributions. First, we crisply articulate why and when observational criteria fail, thus formalizing what was before a matter of opinion. Second, our approach exposes previously ignored subtleties and why they are fundamental to the problem. Finally, we put forward natural causal non-discrimination criteria and develop algorithms that satisfy them.

Motivation & Objective

• Clarify limitations of observational fairness criteria and formalize causal criteria for discrimination.
• Distinguish between protected attributes and their proxies within a causal graph.
• Introduce natural causal criteria based on interventions to describe forms of discrimination.
• Develop and illustrate algorithms for removing proxy discrimination under linear causal models.

Proposed method

• Introduce causal graphs and structural equation models to represent data generating processes involving protected attributes, proxies, features, and predictors.
• Define unresolved discrimination and proxy discrimination with respect to resolving and proxy variables in the graph.
• Propose an intervention-based definition of proxy discrimination using do-calculus and derive conditions for removing it in linear models.
• Provide a procedural guide (and a concrete linear example) to modify predictors to satisfy non-discrimination constraints.
• Discuss relationships to existing fairness notions such as demographic parity and equalized odds, and show how interventional criteria relate to individual fairness concepts.

Experimental results

Research questions

• RQ1 How do observational fairness criteria fail to distinguish certain discriminatory scenarios in causal terms?
• RQ2 How can interventions on proxies or resolving variables formally characterize and prevent discrimination in predictions?
• RQ3 What practical procedures can ensure predictors exhibit no proxy discrimination or no unresolved discrimination under given causal structures?
• RQ4 How do proxy discrimination criteria relate to other fairness notions and to individual fairness?

Key findings

• Observational criteria cannot generally determine whether a predictor exhibits unresolved discrimination, even for Bayes-optimal predictors.
• A causal framework distinguishing resolving variables and proxies exposes subtleties in fair decision making and supports natural discrimination criteria based on interventions.
• There exist procedures to remove proxy discrimination in linear causal models by intervening on proxies and reexpressing the predictor to cancel undesired causal paths.
• Removing proxy discrimination can be achieved while still allowing the predictor to utilize features with potential proxy discrimination, under suitable expressibility assumptions.
• Unresolved discrimination cannot always be eliminated by simple adjustment unless the predictor has access to certain variables, highlighting limits of purely observational adjustments.