[Paper Review] Training individually fair ML models with Sensitive Subspace Robustness

Mikhail Yurochkin, Amanda Bower+1 more|arXiv (Cornell University)|Jun 28, 2019

Explainable Artificial Intelligence (XAI)41 references53 citations

TL;DR

The paper formulates individual fairness as robustness to sensitive input perturbations via distributionally robust optimization and Wasserstein distances, and presents SenSR to train models that are fair to sensitive subspaces, with theoretical guarantees and empirical results.

ABSTRACT

We consider training machine learning models that are fair in the sense that their performance is invariant under certain sensitive perturbations to the inputs. For example, the performance of a resume screening system should be invariant under changes to the gender and/or ethnicity of the applicant. We formalize this notion of algorithmic fairness as a variant of individual fairness and develop a distributionally robust optimization approach to enforce it during training. We also demonstrate the effectiveness of the approach on two ML tasks that are susceptible to gender and racial biases.

Motivation & Objective

• Motivate and formalize individual fairness as robustness to sensitive perturbations in inputs.
• Introduce distributionally robust fairness (DRF) via fair Wasserstein distance.
• Propose and implement Sensitive Subspace Robustness (SenSR) to train individually fair models.
• Provide theoretical guarantees on uniform convergence and fairness certification.
• Demonstrate effectiveness on sentiment analysis and income prediction tasks.

Proposed method

• Define a fair metric d_x that ignores specified sensitive directions; construct a fair Wasserstein distance W using d_z that couples inputs and outputs.
• Audit model fairness by solving a distributionally robust optimization (DRO) problem to maximize loss over distributions within an ε-Wasserstein ball (Equation 2.1).
• Convert the DRO problem to a tractable dual form (Equation 2.3) and solve via stochastic optimization (Algorithm 1).
• Cast fair training as a minimax DRO problem (Equation 2.5) and solve with a SenSR algorithm inspired by adversarial training (Algorithm 2).
• Provide uniform convergence guarantees (Propositions 3.1–3.3) showing conditions under which SenSR yields individually fair models and can certify fairness.
• Relate the approach to existing DRO and adversarial fairness methods and discuss cost function discrepancies (A3) and metric learning.

Experimental results

Research questions

• RQ1 How can individual fairness be enforced when the fair metric is learned from data or observed attributes?
• RQ2 Can distributionally robust optimization with a fair Wasserstein distance yield models that perform well while being robust to sensitive perturbations?
• RQ3 What are the theoretical guarantees (uniform convergence, fairness certification) for SenSR in this setting?
• RQ4 How does SenSR perform on tasks with observed and unobserved sensitive attributes (e.g., sentiment with names, income prediction)?

Key findings

• SenSR/DRF training yields models with substantially reduced gaps on fairness metrics while preserving accuracy (e.g., sentiment task shows large reductions in Race and Gend. gaps).
• Empirical results on sentiment analysis show SenSR and SenSR-E achieve near-equal sentiment predictions across names, outperforming baselines and other debiasing methods in individual fairness.
• In the Adult dataset, SenSR improves S-Con and GR-Con fairness measures and achieves competitive or better group fairness metrics compared to baselines, with a modest drop in balanced accuracy.
• The paper provides a certificate-like generalization bound (Equation 3.3 and Propositions 3.1–3.3) that linking DRO loss to fairness guarantees, enabling fairness certification for practitioners.
• The approach enables learning a fair metric from data (observed or unobserved sensitive attributes) and demonstrates robustness to perturbations in sensitive subspaces across tasks.