An efficient ensemble method for classifying skewed data streams

Abstract

Class distributions of data streams in real application are usually unbalanced, they are hence called Skewed Data Streams (abbreviated as SDS). However, in the classification of SDS, it is a challenge for traditional methods because of the difficulty in the recognition of minority classes. Therefore, many approaches have been proposed to improve the recognition rate of minority classes, while they are time-consuming. Motivated by this, we propose an efficient Ensemble method for Classifying SDS called ECSDS. Our algorithm creates multiple classifiers based on C4.5, and adopts the threshold of F1-value to limit the updating frequency of classifiers. Meanwhile, it adds misclassified positive instances into the training data to guarantee the effectiveness of classifiers when updating. Experimental studies demonstrate that our proposed method enables reducing the time overhead and maintains a good performance on the classification accuracy. © 2012 Springer-Verlag.