Accuracy weighted diversity-based online boosting. (1st December 2020)

• Record Type: Journal Article
• Authors: Baidari, Ishwar; Honnikoll, Nagaraj
• Is Part Of: Expert systems with applications. Volume 160(2020)
• Publication Date: 2020-12-01
• Subjects: Data stream -- Concept drift -- Online boosting -- Diversity; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2020.113723
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: AWDOB: a new ensemble method inspired on ADOB for data stream. A new weighting scheme is defined for the calculation of current expert weight. Experiments were run on 10 real world and 32 artificial data stream datasets. AWDOB presented overall best accuracy result when compared with other methods. Abstract: Target distributional change occurring in a data stream known as concept drift, causes a challenging task for an online learning method, as the accuracy of an online learning method may decrease due to these changes. In this paper, the Accuracy Weighted Diversity-based Online Boosting (AWDOB) method has been proposed, which is based on Adaptable Diversity-based Online Boosting (ADOB) and, other modifications. More precisely, AWDOB uses the proposed accuracy weighting scheme which is based on previous expert's results of the sums of correctly classified and incorrectly classified instances to calculate the weight of current expert, which improved the overall accuracy of the AWDOB. Experiments were conducted to compare the accuracy results of AWDOB against other methods using ten real-world datasets and thirty-two artificial datasets. Artificial datasets were generated by the four artificial data generators which included gradual and abrupt concept drifts within them. Experimental results suggest that AWDOB beats the accuracy results of other tested methods.