FairnessBiasAccountabilitystat.ML
Domain Adaptation meets Individual Fairness. And they get along
Debarghya Mukherjee, Felix Petersen, Mikhail Yurochkin, Yuekai Sun
May 1, 2022
Neural Information Processing Systems
21 citations
1 influential
Key Takeaways
-
Algorithmic bias and performance gaps often stem from the same root cause as other ML failures: The paper demonstrates that distributional shifts—when a model encounters data different from its training set—are a major driver of both poor performance on underrepresented groups and overall model failures, suggesting these problems can be addressed through similar solutions.
-
Fairness interventions can unexpectedly improve model performance across the board: Enforcing individual fairness constraints (rules ensuring similar individuals are treated similarly) actually helps ML models perform better on new, unfamiliar data, even though fairness techniques were not originally designed for this purpose.
-
Domain adaptation techniques offer a new toolkit for reducing algorithmic bias: Methods traditionally used to help models work across different data environments can be repurposed to mitigate bias against demographic groups, opening a new avenue for building fairer AI systems.
-
The connection between fairness and robustness is bidirectional: The paper shows that addressing one problem (fairness or distribution shifts) can help solve the other, suggesting that responsible AI development should consider these challenges together rather than in isolation.
Abstract
Many instances of algorithmic bias are caused by distributional shifts. For example, machine learning (ML) models often perform worse on demographic groups that are underrepresented in the training data. In this paper, we leverage this connection between algorithmic fairness and distribution shifts to show that algorithmic fairness interventions can help ML models overcome distribution shifts, and that domain adaptation methods (for overcoming distribution shifts) can mitigate algorithmic biases. In particular, we show that (i) enforcing suitable notions of individual fairness (IF) can improve the out-of-distribution accuracy of ML models under the covariate shift assumption and that (ii) it is possible to adapt representation alignment methods for domain adaptation to enforce individual fairness. The former is unexpected because IF interventions were not developed with distribution shifts in mind. The latter is also unexpected because representation alignment is not a common approach in the individual fairness literature.