Metric Learning from Imbalanced Data with Generalization Guarantees

Since many machine learning algorithms require a distance metric to capture dis/similarities between data points, metric learning has received much attention during the past decade. Surprisingly, very few methods have focused on learning a metric in an imbalanced scenario where the number of positive examples is much smaller than the negatives, and even fewer derived theoretical guarantees in this setting. Here, we address this difficult task and design a new Mahalanobis metric learning algorithm (IML) which deals with class imbalance. We further prove a generalization bound involving the proportion of positive examples using the uniform stability framework. The empirical study performed on a wide range of datasets shows the efficiency of IML. (C) 2020 Elsevier B.V. All rights reserved.