A Double-Ensemble Approach for Classifying Skewed Data Streams

Abstract

Nowadays, many applications need to handle large amounts of streaming data, which often presents a skewed distribution, i.e. one or more classes are largely under-represented in comparison to the others. Unfortunately, little effort has been directed towards the classification of skewed data streams, although class-imbalance learning has already been studied in the area of pattern recognition on static data. Furthermore, while existing class-imbalance learning methods increase the recognition accuracy on minority class, they often harm the global classification accuracy. Motivated by these observations, we develop an approach suited for classifying skewed data streams, which integrates two ensembles of classifiers, each one suited for non-skewed and skewed data. This approach substantially increases the global accuracy compared to existing classification methods for skewed data. Experimental tests have been carried out on three public datasets showing interesting results. As a further contribution, we will study metrics to evaluate the performance of skewed data streams classification. We will also review the literature on class-imbalance learning, and skewed data streams classification.