Drift Adaptation via Joint Distribution Alignment

Abstract

Machine learning in evolving environment faces challenges due to concept drift. Most concept drift adaptation methods focus on modifying the model. In this paper, a method, Drift Adaptation via Joint Distribution Alignment (DAJDA), is proposed. DAJDA performs a linear transformation to the drift instances instead of modifying model. Instances are transformed into a common feature space, reducing the discrepancy of distributions before and after drift. Experimental studies show that DAJDA has abilities to improve the performance of learning model under concept drift.