Title:

 

Using Artificial Intelligence in critical decision-making applications introduces deep risks regarding transparency, fairness, and accountability. We know that embedded biases in commonly used training data often distribute predictive errors unevenly across individuals and subgroups — a problem severely compounded by the opacity of modern AI pipelines. Explainability methods, particularly counterfactual explanations, offer a robust framework to understand these fairness implications. Succinctly, a counterfactual explanation encodes a minimal set of input modifications required to alter a model’s prediction. It thus provides a reliable mechanism to compare instances and evaluate equitable outcomes. Here we present novel techniques for deriving local counterfactual explanations by optimizing for feasibility, simplicity, and compactness. Furthermore, to mitigate the scalability limitations of instance-level algorithms, we extend our scope to consider global counterfactual frameworks. Our methods synthesize succinct, actionable recourse strategies that deliver scalable insights applicable across large population subgroups.