The Limitations of Model Retraining in the Face of Performativity

TL;DR

This paper investigates the challenges of retraining machine learning models under performative shifts, showing that naive approaches are suboptimal and proposing regularization as a solution to achieve optimality.

Contribution

It introduces a theoretical analysis of retraining under performative shifts and demonstrates that regularization can provably improve model performance in such settings.

Findings

Naive retraining can be suboptimal under performative shifts.

Regularization improves retraining outcomes in performative environments.

Theoretical guarantees show regularization attains optimal models.

Abstract

We study stochastic optimization in the context of performative shifts, where the data distribution changes in response to the deployed model. We demonstrate that naive retraining can be provably suboptimal even for simple distribution shifts. The issue worsens when models are retrained given a finite number of samples at each retraining step. We show that adding regularization to retraining corrects both of these issues, attaining provably optimal models in the face of distribution shifts. Our work advocates rethinking how machine learning models are retrained in the presence of performative effects.