Concept drift visualization

Mining data stream are facing many challenges now, one of them is concept drift problem. In many practical applications, concept drift usually affects the classification performance for data stream, or even make the classifier failed. However, most of the proposed methods are mainly focusing on solving concept drift from the data value point of view, and very little attention has been focused on mining the knowledge in the data concept level. Motivated by this, in this paper, we use Kullback-Leibler divergence (KL-divergence) algorithm to detect concept drift dynamically. Meanwhile, we also construct a concept pool to reserve distinct concepts in data stream and analyze the concept transformation information. Experimental studies on two real-world data sets demonstrate that the proposed concept visualization method and concept transformation map could effectively and efficiently mine concept drifts relationship from the noisy streaming data. Copyright © 2013 Binary Information Press.