[Paper Review] Fair Predictors under Distribution Shift

Harvineet Singh, Rina Singh+2 more|arXiv (Cornell University)|Nov 2, 2019

Explainable Artificial Intelligence (XAI)22 references4 citations

TL;DR

This paper proposes a method to train machine learning predictors that remain both accurate and fair under distribution shifts by selecting features under fairness constraints, leveraging causal domain adaptation. It achieves stable performance in prediction and fairness across unseen target distributions, demonstrated on synthetic data and a real-world acute kidney injury diagnosis task under measurement policy shift and selection bias.

ABSTRACT

Recent work on fair machine learning adds to a growing set of algorithmic safeguards required for deployment in high societal impact areas. A fundamental concern with model deployment is to guarantee stable performance under changes in data distribution. Extensive work in domain adaptation addresses this concern, albeit with the notion of stability limited to that of predictive performance. We provide conditions under which a stable model both in terms of prediction and fairness performance can be trained. Building on the problem setup of causal domain adaptation, we select a subset of features for training predictors with fairness constraints such that risk with respect to an unseen target data distribution is minimized. Advantages of the approach are demonstrated on synthetic datasets and on the task of diagnosing acute kidney injury in a real-world dataset under an instance of measurement policy shift and selection bias.

Motivation & Objective

• Address the challenge of maintaining both predictive accuracy and fairness when machine learning models face distribution shifts in real-world deployments.
• Extend domain adaptation techniques to include fairness as a stability criterion, not just predictive performance.
• Develop a framework that minimizes risk under unseen target data distributions while enforcing fairness constraints.
• Identify and select a subset of features that enable stable performance in both prediction and fairness metrics across distribution shifts.

Proposed method

• Formulate the problem within a causal domain adaptation framework to model the relationship between source and target distributions.
• Select a subset of features for training that optimizes for both predictive risk and fairness constraints under the target distribution.
• Use risk minimization objectives that account for both prediction error and fairness violations in the target distribution.
• Incorporate fairness constraints into the training process to ensure that model performance remains stable across distribution shifts.
• Apply the method to synthetic data and a real-world acute kidney injury diagnosis dataset under measurement policy shift and selection bias.
• Leverage causal assumptions to identify features that are robust to distribution shifts while preserving fairness.

Experimental results

Research questions

• RQ1 Can a machine learning model maintain both high predictive accuracy and fairness under distribution shifts?
• RQ2 How can fairness be stabilized across different data distributions without compromising predictive performance?
• RQ3 Which features are most effective for training predictors that generalize well in terms of both accuracy and fairness under distribution shift?
• RQ4 To what extent does the proposed method reduce fairness degradation under measurement policy shift and selection bias in real-world data?

Key findings

• The proposed method achieves stable predictive and fairness performance across distribution shifts, outperforming baseline approaches in both synthetic and real-world settings.
• Feature selection under fairness constraints significantly improves robustness to distribution shifts, particularly under measurement policy shift.
• The method reduces fairness degradation in the target distribution compared to models trained without fairness-aware feature selection.
• Empirical results on the acute kidney injury dataset show improved generalization in both prediction accuracy and fairness metrics under selection bias.
• The approach demonstrates that fairness can be preserved as a stability criterion alongside predictive performance in domain adaptation.
• Causal modeling of domain shift enables identification of features that are both predictive and fair under distributional changes.