Bilkent University
Department of Computer Engineering


Geometrically Optimum and Online Weighted Ensemble Classifier (GOOWE) on Evolving Data Streams


Hamed Rezanejad Asl- Bonab
MS Student
Computer Engineering Department
Bilkent University

Designing adaptive classifiers for evolving stream data is a challenging task due to the size and changing nature of data streams. Among existing classifiers, ensemble-based approach is one of the best solutions to this problem. However, there are a few data-fusion algorithms for combining the scores of classifiers in an ensemble. This becomes more important in evolving environments because in different situations it is possible that a subset of classifiers in the ensemble outperforms others. Our aim is to introduce a novel geometric framework for weighting and combining classifiers in an ensemble. In this framework, we propose a dynamic weighting approach for classifiers called “Geometrically Optimum and Online Weighted Ensemble Classifier (GOOWE)” based on the Euclidean distance between scores and ideal points. In order to measure the robustness of proposed framework, we used real-world datasets and synthetic data generators using MOA Libraries. Besides, we discuss experimental setup and finally analyze simulation results. For the sake of comparisons, we include eight latest online ensemble methods and statistical tests indicate a significant improvement in accuracy.


DATE: 16 November, 2015, Monday @ 15:40