Suitability Filter: A Statistical Framework for Classifier Evaluation in Real-World Deployment Settings

TL;DR

The paper introduces the suitability filter, a statistical framework that detects classifier performance degradation on unlabeled user data by analyzing suitability signals and employing hypothesis testing, enhancing deployment reliability in safety-critical domains.

Contribution

It presents a novel, modular framework that uses suitability signals and statistical testing to evaluate classifier performance without ground truth labels, addressing covariate shift challenges.

Findings

Reliably detects performance drops due to covariate shift.

Effective across various classification tasks and domains.

Enables proactive mitigation of model failures.

Abstract

Deploying machine learning models in safety-critical domains poses a key challenge: ensuring reliable model performance on downstream user data without access to ground truth labels for direct validation. We propose the suitability filter, a novel framework designed to detect performance deterioration by utilizing suitability signals -- model output features that are sensitive to covariate shifts and indicative of potential prediction errors. The suitability filter evaluates whether classifier accuracy on unlabeled user data shows significant degradation compared to the accuracy measured on the labeled test dataset. Specifically, it ensures that this degradation does not exceed a pre-specified margin, which represents the maximum acceptable drop in accuracy. To achieve reliable performance evaluation, we aggregate suitability signals for both test and user data and compare these empirical distributions using statistical hypothesis testing, thus providing insights into decision uncertainty. Our modular method adapts to various models and domains. Empirical evaluations across different classification tasks demonstrate that the suitability filter reliably detects performance deviations due to covariate shift. This enables proactive mitigation of potential failures in high-stakes applications.