A model can produce statistically fair predictions across demographic groups while the explanations it gives for those predictions — via methods like SHAP or LIME — are inconsistent or lower-quality for some groups than others. This project audits that gap: it measures whether post-hoc explanation methods themselves exhibit disparity, independent of whether the underlying model’s outputs are fair.

Beyond measuring the problem, the work proposes mitigation strategies so that explanation quality (not just prediction accuracy) can be held to a fairness standard — a distinction that matters for any setting where a model’s stated reasoning, not just its answer, needs to be trusted equally across users.

The full abstract is available via IEEE Xplore; the paper was not yet publicly indexed with an open abstract at the time this page was written.