Matched-Pair Machine Learning

Abstract

Following an analogous distinction in statistical hypothesis testing and motivated by chemical plume detection in hyperspectral imagery, we investigate machine-learning algorithms where the training set is comprised of matched pairs. We find that even conventional classifiers exhibit improved performance when the input data have a matched-pair structure, and we develop an example of a “dipole” algorithm to directly exploit this structured input. In some scenarios, matched pairs can be generated from independent samples, with the effect of not only doubling the nominal size of the training set, but of providing the matched-pair structure that leads to better learning. The creation of matched pairs from a dataset of interest also permits a kind of transductive learning, which is found for the plume detection problem to exhibit improved performance. Supplementary materials for this article are available online.