[Paper Review] Robust Fairness-aware Learning Under Sample Selection Bias

Wei Du, Xintao Wu|arXiv (Cornell University)|May 24, 2021

Ethics and Social Impacts of AI24 references4 citations

TL;DR

This paper proposes a robust and fair learning framework to address sample selection bias in machine learning, combining reweighing for bias correction and minimax optimization to ensure fairness and accuracy under distributional shifts. The method achieves improved prediction performance and fairness on test data, even when test data is unavailable, by modeling uncertainty via Wasserstein balls and enforcing fairness under worst-case conditions.

ABSTRACT

The underlying assumption of many machine learning algorithms is that the training data and test data are drawn from the same distributions. However, the assumption is often violated in real world due to the sample selection bias between the training and test data. Previous research works focus on reweighing biased training data to match the test data and then building classification models on the reweighed training data. However, how to achieve fairness in the built classification models is under-explored. In this paper, we propose a framework for robust and fair learning under sample selection bias. Our framework adopts the reweighing estimation approach for bias correction and the minimax robust estimation approach for achieving robustness on prediction accuracy. Moreover, during the minimax optimization, the fairness is achieved under the worst case, which guarantees the model's fairness on test data. We further develop two algorithms to handle sample selection bias when test data is both available and unavailable. We conduct experiments on two real-world datasets and the experimental results demonstrate its effectiveness in terms of both utility and fairness metrics.

Motivation & Objective

• To address the challenge of sample selection bias in machine learning, where training and test data distributions differ, leading to degraded model performance.
• To ensure fairness in classification models under such distributional shifts, particularly when biased training data may lead to unfair predictions on test data.
• To develop a unified framework that simultaneously corrects for sample selection bias and enforces fairness under worst-case distributional uncertainty.
• To provide practical algorithms for both scenarios—when test data is available and when it is not—offering robustness and fairness guarantees.
• To demonstrate the effectiveness of the approach through empirical evaluation on real-world datasets, showing improvements in both utility and fairness metrics.

Proposed method

• Uses reweighing estimation to correct sample selection bias by adjusting training example weights based on the ratio of selection probabilities between training and test data.
• Applies minimax robust estimation over a Wasserstein ball centered at the empirical training distribution to ensure worst-case robustness in prediction accuracy.
• Incorporates fairness constraints into the loss function via boundary fairness approximation, penalizing deviations in risk difference between protected and unprotected groups.
• For RFLearn 1 (test data available), estimates selection probabilities using density ratio estimation between training and test data.
• For RFLearn 2 (test data unavailable), assumes cluster-wise uniform selection probabilities and robustifies estimation within a Wasserstein ball for each cluster.
• Optimizes the modified loss function under worst-case distributions to guarantee fairness and accuracy robustness on unseen test data.

Experimental results

Research questions

• RQ1 How can we correct sample selection bias in training data while ensuring fairness in the resulting classifier on test data?
• RQ2 Can a minimax robust optimization framework be effectively combined with fairness constraints to improve model reliability under distributional shift?
• RQ3 How can we achieve robust and fair learning when test data is not available for selection probability estimation?
• RQ4 What is the impact of using Wasserstein ambiguity sets on fairness and accuracy under sample selection bias?
• RQ5 How does the proposed framework compare to existing methods in terms of fairness and utility metrics on real-world datasets?

Key findings

• The proposed framework significantly improves prediction accuracy and fairness on test data under sample selection bias, outperforming baseline methods in both utility and fairness metrics.
• RFLearn 1, which uses test data for density ratio estimation, achieves better fairness and accuracy than standard reweighing and fairness-aware baselines.
• RFLearn 2, designed for the case when test data is unavailable, maintains strong performance by leveraging cluster-based selection probability assumptions and robust optimization.
• The use of a Wasserstein ball for modeling distributional uncertainty leads to more robust predictions under worst-case distribution shifts.
• Incorporating fairness constraints during minimax optimization ensures that fairness is preserved even when the test distribution deviates from the training distribution.
• Empirical results on two real-world datasets confirm that the framework effectively balances robustness and fairness, with measurable gains in risk difference and accuracy.