Actionable Recourse

A counterfactual explanation tells someone the nearest input that would have flipped a model's decision — for example, a loan applicant would have been approved with $4,000 more in reported income. Actionable recourse tightens that idea with a practical constraint: the counterfactual has to describe a change the person could actually make. Raising your income by $4,000 overnight isn't something most applicants can do; lowering your age by five years or changing your zip code isn't something anyone can do at all. A recourse recommendation restricts the search to features the person can plausibly act on — pay down a credit card balance, wait six months to build tenure — and ignores or fixes immutable ones like age, sex, or birthplace.

This distinction matters for both fairness and compliance: an explanation that quietly suggests "be five years younger" is useless and can even be discriminatory to surface, whereas "reduce your credit utilization below 30%" is something a person can go do. Building a recourse system means specifying, feature by feature, which variables are actionable, which direction they can move (income can go up but rarely down at will), and how costly each change is, then searching for the cheapest actionable change that flips the model's decision.