New Management Operations on Classifiers Pool to Track Recurring Concepts

Abstract

Handling recurring concepts has become of interest as a challenging problem in the field of data stream classification in recent years.  One main feature of data streams is that they appear in nonstationary environments. This means that the concept which the data are drawn from, changes over the time. If after a long enough time, the concept reverts to one of the previous concepts, it is said that recurring concepts has occurred. One solution to this challenge is to maintain a pool of classifiers, each representing a concept in the stream. This paper follows this approach and holds an ensemble of classifiers for each concept. As for each received batch of data, a new classifier is created; there will be a huge amount of classifiers which could not be maintained in the pool. To handle the memory limitations, a maximum number of concepts and classifiers are assumed. So the necessity of managing the concepts and classifiers is obvious. This paper presents a novel algorithm to manage the pool. Some pool management operations including merging and splitting the concepts are introduced. Experimental results show the performance dominance of using our method to the most promising stream classification algorithms.