Invariance assumptions for class distribution estimation

TL;DR

This paper investigates how invariance assumptions like covariate shift and joint shift can improve the estimation of class priors under dataset shift, where only features are observed in test data.

Contribution

It analyzes different invariance assumptions and their implications for accurately estimating class distributions in the presence of dataset shift.

Findings

Covariate shift assumption simplifies class prior estimation.

Factorizable joint shift provides a structured approach.

Sparse joint shift offers a flexible model for distribution changes.

Abstract

We study the problem of class distribution estimation under dataset shift. On the training dataset, both features and class labels are observed while on the test dataset only the features can be observed. The task then is the estimation of the distribution of the class labels, i.e. the estimation of the class prior probabilities, in the test dataset. Assumptions of invariance between the training joint distribution of features and labels and the test distribution can considerably facilitate this task. We discuss the assumptions of covariate shift, factorizable joint shift, and sparse joint shift and their implications for class distribution estimation.