Title: Finding the Right Regularization for Performative Learning

Abstract: Performative learning introduces a unique complexity compared to traditional supervised learning: the data distribution shifts in response to the deployed model. This phenomenon often arises when strategic users modify their features to manipulate outcomes, creating a dynamic environment that requires more than just optimizing for the current dataset. Models must account for the fact that their presence may steer the underlying distribution in unforeseen directions, even when the specifics of such shifts remain unknown. This study investigates the role of regularization in mitigating performative effects, specifically analyzing its impact within high-dimensional ridge regression. Our findings reveal a dichotomy: while performative dynamics tend to degrade test risk in standard population settings, they can actually be advantageous in over-parameterized scenarios where the feature count surpasses the number of samples. Crucially, we demonstrate that the ideal regularization strength should scale proportionally with the magnitude of the performative effect, allowing practitioners to calibrate regularization parameters in advance to counteract these influences. We validate this theoretical insight through empirical tests of optimal regularization settings on both synthetic and real-world data sets.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC