THE ROLE OF CAUSALITY IN ALGORITHMIC RECOURSE

In high-stakes settings like lending or hiring, algorithmic recourse provides individuals with actionable advice to change their outcomes. However, standard recourse methods often ignore whether these actions lead to genuine improvement in an individual's qualifications or merely enable "gaming" of the classifier. This paper investigates how to design recourse policies that promote meaningful improvement while maintaining predictive accuracy, specifically by accounting for the causal relationships between features.

The authors propose a causal performative framework that models how recourse interventions propagate through a structural causal model (SCM). By representing feature dependencies as a directed acyclic graph (DAG), the model captures how changing one feature (e.g., debt) indirectly affects others (e.g., credit score) and the true label (e.g., repayment ability). They formulate the learner's problem as a performative prediction task, where the deployed model influences agent behavior, which in turn shifts the data distribution. The authors provide iterative algorithms—Repeated Risk Minimization (RRM) and Repeated Gradient Descent (RGD)—to compute performatively stable solutions, where the model remains consistent even after agents adapt to it.

The study demonstrates that ignoring causal structure leads to misaligned behavioral responses, where agents manipulate non-causal features to "game" the system, eventually degrading the model's predictive accuracy. In contrast, causal recourse accounts for these interdependencies, leading to stable equilibria that reduce the incentive for strategic manipulation. Theoretical analysis establishes conditions under which these algorithms converge to stable points, and empirical experiments on both semi-synthetic and real-world credit datasets confirm that the causal approach outperforms standard empirical risk minimization by reducing the need for frequent model retraining.