A Learning Based Hypothesis Test for Harmful Covariate Shift
This addresses the critical need for safe deployment in high-risk domains by enabling timely model retraining, though it is an incremental improvement on existing detection methods.
The paper tackles the problem of detecting harmful covariate shift in machine learning systems by defining it as distributional changes that weaken model generalization, and it shows state-of-the-art performance across various high-dimensional datasets, especially with small test samples.
The ability to quickly and accurately identify covariate shift at test time is a critical and often overlooked component of safe machine learning systems deployed in high-risk domains. While methods exist for detecting when predictions should not be made on out-of-distribution test examples, identifying distributional level differences between training and test time can help determine when a model should be removed from the deployment setting and retrained. In this work, we define harmful covariate shift (HCS) as a change in distribution that may weaken the generalization of a predictive model. To detect HCS, we use the discordance between an ensemble of classifiers trained to agree on training data and disagree on test data. We derive a loss function for training this ensemble and show that the disagreement rate and entropy represent powerful discriminative statistics for HCS. Empirically, we demonstrate the ability of our method to detect harmful covariate shift with statistical certainty on a variety of high-dimensional datasets. Across numerous domains and modalities, we show state-of-the-art performance compared to existing methods, particularly when the number of observed test samples is small.