Thread by @h_nilforoshan

Excited and honored to share that our #ICML2022 paper, “Causal Conceptions of Fairness and their Consequences” (w/ @jgaeb1, Ravi Shroff, and @5harad) has received an outstanding paper award!!! See 🧵for key findings and results

Paper: arxiv.org/abs/2207.05302 1/17

@jgaeb1 @5harad We find that, surprisingly, the paradigm shift in fair ML towards causal reasoning is *harmful* to common policy goals, including diversity and equity efforts. 2/17

Our results are counterintuitive because causal approaches may seem like policy *improvements* over traditional, non-causal fairness criteria. For instance, a traditional approach in fair ML (blinding) is to simply remove features like race or gender from decisions. 3/17

But *indirect* bias could persist, e.g., if earlier discrimination against historically disadvantaged groups causes differences in other features—e.g., standardized test scores in college admissions, income in credit scoring, etc. 4/17

This is why one causal approach is to reduce all unfair causal effects. For example, counterfactual fairness requires that protected characteristics (e.g. race, gender) have no effect on decisions—even indirectly through features like standardized test score or income. 5/17

(To engage with the literature, we assume it’s possible to conceptualize a “causal effect” of immutable attributes like race or gender. However, counterfactuals of immutable attributes raise their own set of thorny philosophical and practical challenges.) 6/17

Surprisingly, this is a *very* strong constraint: under mild assumptions, we prove that counterfactual fairness—rather than just mitigating bias, as one might expect—instead requires *throwing features away entirely* when they exhibit *any* amount of direct or indirect bias. 7/17

Thus, in real-world settings involving race, gender, etc. the only “counterfactually fair” decision policies are likely to be lotteries (i.e., throwing away all features) or decisions based on a small number of unrelated features, like age. 8/17

Counterfactual fairness is extreme, but we find that this basic story is true more generally for prominent causal fairness definitions… 9/17

We survey the literature and identify two approaches to causal fairness.

1. Reducing the effect of protected attributes on decisions (e.g. counterfactual fairness)
2. Reducing the effect of decisions on counterfactual disparities (e.g. counterfactual equalized odds) 10/17

We find that both approaches of fairness definitions lead to suboptimal policies, perversely harming the groups they were designed to protect. 11/17

For example, in college admissions, the admissions committee might seek to maximize both the number of students who attain degrees and diversity, by recruiting students from a historically disadvantaged group. 12/17

This plot from our simulation compares the outcomes of different admissions policies. The pink Pareto frontier shows policies that directly optimize for desired outcomes—i.e., that maximize diversity (X-axis) and degree attainment (Y-axis). 13/17

Perversely, one can admit more applicants from the target group *and* graduate more applicants “for free” with a policy that ignores these definitions and directly optimizes for desired outcomes like diversity. 14/17

The results of our simulation hold in great generality. For natural families of utility functions, we prove that both approaches almost always—in a measure theoretic sense—lead to (strongly) Pareto dominated policies, i.e., policies beneath the frontier. 15/17

What role, then, should causality play in algorithmic fairness? We argue for an alternative approach: we should encode preferences for diversity in the utility (or loss) itself. Then, we need causality to find the policies with the most desirable expected outcomes. 16/17

We hope this work illuminates the field of causal fairness and its potential adverse consequences.

Come to our talk at #ICML2022 on Thursday!!!

Paper: arxiv.org/abs/2207.05302
Slides: hamedn.com/cf.pdf
Code: github.com/stanford-polic…
Theory: jgaeb.com/2022/07/18/pre… 17/17