On the Robustness of Fairness Practices: A Causal Framework for Systematic Evaluation

TL;DR

The paper tackles the problem that established fairness practices for ML software may not generalize under data shifts or labeling faults. It introduces a causal-graph–based robustness-testing framework that uses CPDAG-equivalence classes and structural causal models to generate neighboring data distributions, then evaluates how pre-/in-/post-processing fairness interventions perform across these variations. Across six real-world datasets and multiple learning algorithms, robustness of fairness practices is shown to be highly context-dependent on the underlying causal structure, with some feature-selection methods proving relatively robust while many post-processing mitigations fail to be robust in many cases. The work provides a practical tool and methodology for SE practitioners to stress-test fairness interventions before deployment and argues for treating fairness as a robustness concern in software engineering, with implications for future causal-theory–driven fairness testing.

Abstract

Machine learning (ML) algorithms are increasingly deployed to make critical decisions in socioeconomic applications such as finance, criminal justice, and autonomous driving. However, due to their data-driven and pattern-seeking nature, ML algorithms may develop decision logic that disproportionately distributes opportunities, benefits, resources, or information among different population groups, potentially harming marginalized communities. In response to such fairness concerns, the software engineering and ML communities have made significant efforts to establish the best practices for creating fair ML software. These include fairness interventions for training ML models, such as including sensitive features, selecting non-sensitive attributes, and applying bias mitigators. But how reliably can software professionals tasked with developing data-driven systems depend on these recommendations? And how well do these practices generalize in the presence of faulty labels, missing data, or distribution shifts? These questions form the core theme of this paper.

Figures (8)

• Figure 1: Causal Framework for Robust Fairness.
• Figure 2: Causal graph (a) generated by GES algorithm with unresolved edges. Causal graphs (b) and (c) are two equivalent DAGs.
• Figure 3: Result of EOD on the causal graph \ref{['fig:Adult_DAGs']} (b).
• Figure 4: Result of EOD on the causal graph \ref{['fig:Adult_DAGs']} (c).
• Figure 5: Selection Operators.