Algorithmic Recourse

Algorithmic recourse asks what a person could feasibly change to receive a different outcome from an automated decision, and whether that recommendation is useful and fair. Methods include counterfactual optimization and search, causal models that represent how actions affect features and outcomes, and rule-based or tree-based summaries of shared options. Evaluations use settings such as credit, hiring, and clinical risk, testing whether recommendations are valid, affordable, plausible, robust, and achievable under real constraints. A central tension is that a change that flips a model’s prediction may be costly or impossible for the person, and widespread adoption can shift thresholds or reinforce existing disparities.