An Enhanced Online Boosting Ensemble Classification Technique to Deal with Data Drift

Abstract

Over the last two decades, big data analytics has become a requirement in the research industry. Stream data mining is essential in many areas because data is generated in the form of streams in a wide variety of online applications. Along with the size and speed of the data stream, concept drift is a difficult issue to handle. This paper proposes an Enhanced Boosting-like Online Learning Ensemble Method based on a heuristic modification to the Boosting-like Online Learning Ensemble (BOLE). This algorithm has been improved by implementing a data instance that retains the previous state policy. During the boosting phase of this modified algorithm, the selection and voting strategy for an instance is advanced. Extensive experimental results on a variety of real-world and synthetic datasets show that the proposed method adequately addresses the drift detection problem. It has outperformed several state-of the art boosting-based ensembles dedicated to data stream mining (statistically). The proposed method improved overall accuracy by 1.30 percent to 14.45 percent when compared to other boosting-based ensembles on concept drifted datasets.