This article presents research on how to improve the accuracy of predictions in Machine Learning when there is a covariate shift – a change in the distribution of data between training and testing. The authors propose a method of calculating ‘importance’ that takes into account information about the targets, not just the covariates. This approach is shown to improve the performance of the Kullback-Leibler Importance Estimation Procedure (KLIEP) and makes it possible to apply the method to a real-world plankton classification problem. The research concludes that error estimation is more accurate when target information is used.
Publication date: 2 Feb 2024
Project Page: http://www.aic.uniovi.es
Paper: https://arxiv.org/pdf/2402.01450
