Abstract: Machine learning systems are widely deployed, including for decision making processes that affect people’s lives, such as loan approval or hiring. Gaming the system by modifying one’s features therefore creates an iterative feedback loop, where the data distribution changes under the performative effect of the model. Performative learning tackles this setting and aims to optimize not a single time step model for a fixed distribution, but to find an optimal long term solution. In this talk, we will introduce performative learning and present two results for optimizing performative risk.

Refs: https://arxiv.org/abs/2411.02023 (NeurIPS 2024) and https://arxiv.org/abs/2510.12249 (ICML 2026)