Answers are grounded in this paper and cite their sources.
Automated credit scoring increasingly relies on black‐box models, raising concerns about transparency and the right to algorithmic recourse under the EU AI Act and GDPR. Counterfactual explanations answer ‘what should change for a different outcome?’, yet existing methods often change immutable attributes such as age or gender, or ignore the heterogeneous economic costs of feature changes. This paper introduces CA‐CF (Constrained Actionable Counterfactual Explanations), a gradient‐based framework that jointly optimizes validity, proximity, economic cost and plausibility. CA‐CF projects each optimization step onto the feasible set defined by an immutable/hard‐actionable/soft‐actionable taxonomy. We prove that this projection is exact, that it rules out constraint violations at every iteration, that the projected subgradient method converges to critical points of the non‐smooth objective and that a preventive gradient mask is redundant. The cost combines and penalties, with weights elicited from experts or learned from data as a differentiable percentile cost. We evaluate CA‐CF on four datasets (German Credit, Give Me Some Credit, Credit Risk and Adult Census) against MIP‐ and DiCE‐based baselines and nine further methods in the CARLA library (308,574 counterfactuals, seed–fold protocol). Projection‐based variants achieve zero immutable and hard‐actionable violations, and CA‐CF delivers recourse at up to lower economic cost than DiCE‐random with a median latency of 11–34 ms. In CARLA, it is one of only two methods that are always successful and always feasible. A confidence margin cuts recourse invalidated by retraining from up to to at most . A post hoc sparsification step cuts the number of changed features by up to at no loss of validity. We report the limits: recourse for tree ensembles is sound and, with query‐based repair, valid in of cases rather than guaranteed and plans lie close to the decision boundary.
No takes yet. Share an insight, caveat, or question.
Ikermane et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: