Prediction-Powered Risk Monitoring of Deployed Models for Detecting Harmful Distribution Shifts

TL;DR

This paper introduces a semi-supervised risk monitoring method called PPRM that detects harmful distribution shifts in deployed models by combining synthetic and true labels, providing finite-sample guarantees.

Contribution

It proposes a novel prediction-powered risk monitoring approach that offers assumption-free, finite-sample guarantees for detecting harmful shifts in dynamic environments.

Findings

Effective detection of harmful shifts in image classification, LLM, and telecommunications tasks.

Provides finite-sample guarantees with low false alarm rates.

Outperforms existing methods in various monitoring scenarios.

Abstract

We study the problem of monitoring model performance in dynamic environments where labeled data are limited. To this end, we propose prediction-powered risk monitoring (PPRM), a semi-supervised risk-monitoring approach based on prediction-powered inference (PPI). PPRM constructs anytime-valid lower bounds on the running risk by combining synthetic labels with a small set of true labels. Harmful shifts are detected via a threshold-based comparison with an upper bound on the nominal risk, satisfying assumption-free finite-sample guarantees in the probability of false alarm. We demonstrate the effectiveness of PPRM through extensive experiments on image classification, large language model (LLM), and telecommunications monitoring tasks.