A novel weighting method for online ensemble learning with the presence of concept drift

Publication Type:
Conference Proceeding
Citation:
Decision Making and Soft Computing - Proceedings of the 11th International FLINS Conference, FLINS 2014, 2014, pp. 550 - 555
Issue Date:
2014-01-01
Metrics:
Full metadata record
Files in This Item:
Filename Description Size
ThumbnailACCEPTED MANUSCRIPT.pdfAccepted Manuscript version238.05 kB
Adobe PDF
© 2014 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. Ensemble of classifiers is a very popular method for online and incremental learning in non-stationary environment, as it improves the accuracy of single classifiers and is able to recover from drifting concept without explicit drift detection. However, current ensemble weighing methods do not consider the relationship between a test instance and each ensemble member's training domain. As a result, a locally correct ensemble member may be reduced weight unfairly because that its prediction result of an out of domain test instance is wrong. These inaccuracies will increases when there is a significant concept change. In this paper, therefore, we proposed a fuzzy online ensemble weighting method which takes the consideration of the degree of membership of each instance in each ensemble member and a modified majority voting method to improve the ability of ensembles on handling online classification tasks with concept drift.
Please use this identifier to cite or link to this item: