Signal The Role of Causality in Algorithmic Recourse

Summary

The paper examines algorithmic recourse, the practice of providing individuals with actionable changes to improve predicted outcomes in high-stakes classification settings such as loan and mortgage applications. It notes that most existing approaches focus only on flipping a model's prediction without verifying whether recommended changes reflect genuine improvement in a person's qualifications, which can allow strategic gaming of the classifier. The authors formalize this failure mode through a causal performative framework that models how recourse actions propagate through a structural causal model, capturing interactions among features and their effect on the true outcome label. They characterize conditions under which stable solutions exist and can be computed through simple iterative dynamics, showing that recourse policies ignoring causal structure can induce large, misaligned behavioral responses. Experiments on semi-synthetic and real credit datasets found that the causally-grounded approach consistently outperformed standard empirical risk minimization. The approach was also found to reduce the need for repeated model retraining to accommodate distribution shifts caused by strategic behavior.