Estimating and Explaining Model Performance When Both Covariates and Labels Shift

Abstract

Deployed machine learning (ML) models often encounter new user data that differs from their training data. Therefore, estimating how well a given model might perform on the new data is an important step toward reliable ML applications. This is very challenging, however, as the data distribution can change in flexible ways, and we may not have any labels on the new data, which is often the case in monitoring settings. In this paper, we propose a new distribution shift model, Sparse Joint Shift (SJS), which considers the joint shift of both labels and a few features. This unifies and generalizes several existing shift models including label shift and sparse covariate shift, where only marginal feature or label distribution shifts are considered. We describe mathematical conditions under which SJS is identifiable. We further propose SEES, an algorithmic framework to characterize the distribution shift under SJS and to estimate a model's performance on new data without any labels. We conduct extensive experiments on several real-world datasets with various ML models. Across different datasets and distribution shifts, SEES achieves significant (up to an order of magnitude) shift estimation error improvements over existing approaches.Preprint. Under review.

Factorizable Joint Shift in Multinomial Classification

“…Assumption 3.1, complemented by Assumption 3.7 below, can be reconciled with the setting of Chen et al (2022) in the following way:…”

“…The notion of Sparse Joint Shift (SJS) was introduced by Chen et al (2022) as a tractable model of dataset shift "which considers the joint shift of both labels and a few features". In this paper, we reanalyse the notion in some depth, looking closer at its connection to prior probability shift and the link between SJS and covariate shift.…”

“…We revisit the notion of SJS in a probabilistic measure theory setting that allows for the easy deployment of powerful mathematical tools like the tower property of conditional expectation or Dynkin's theorem which are less readily available within the density-focussed approach used by Chen et al (2022). This paper can be read as a commentary on Chen et al (2022) but also as a self-contained presentation of SJS.…”

“…• We comment on the approaches to the estimation of importance weights proposed by Chen et al (2022) and identify certain issues that might affect their performance (Section 5 below).…”

“…Properties of SJS are discussed in some detail in Section 4. In Section 5, two approaches by Chen et al (2022) to the estimation of densities under the assumption of SJS are presented and commented on.…”