Distributed Online Ensemble Learning Based on Orthogonal Transformation

Abstract

A new distributed online learning scheme for classifying data captured from distributed data sources is proposed in this paper. The scheme consists of multiple distributed learners that independently classify different streams of data. Each local learner uses an ensemble classifier trained by shared data to make a prediction. We propose a novel form of shared data, that is, the covariance matrix and mean vector, that has small and stable network traffic when transmitted between nodes. Then, we provide a systematic online ensemble learning approach based on these shared data. In contrast to boosting and bagging, our proposed learning approach is based on orthogonal transformation, which can increase the differences between individual learners without a significant loss in accuracy. Moreover, we discuss the ensemble maintenance method based on weight to adapt the underlying data dynamics. Empirical studies demonstrate the effectiveness of our approach in comparison to existing state-of-the-art methods on several datasets.

Keywords

• Distributed online learning
• ensemble learning
• covariance matrix
• mean vector
• orthogonal transformation