A Novel Weight Adjustment Method for Handling Concept-Drift in Data Stream Classification

Evolving fuzzy rule-based systems are very powerful methods for online classification of data streams. In these systems, the classifier is updated by generating, removing and modifying fuzzy classification rules. One of the simplest but efficient algorithms of this type is evolving classifier (eClass) that generates fuzzy classification rules without any prior knowledge. However, this algorithm cannot cope properly with drift and shift in the concept of data. In order to improve the performance of this algorithm, in this paper, we propose a scheme that assigns a weight to each fuzzy rule and propose a new efficient online method to adjust the weight of fuzzy classification rules. Using the proposed rule-weighting algorithm, the classifier can quickly cope with drift and shift in the concept of data. Our algorithm is in fact the modified version of a batch mode rule-weight learning algorithm proposed in the past to be consistent with characteristics of data streams. We use some real life and some synthetic datasets to assess the performance of our algorithm in comparison with eClass and some other methods proposed in the past for handling data streams. The results of experiments show that our proposed algorithm performs significantly better than eClass and other methods proposed in the past for classification of data streams.