Classification of Recurrent Concepts with Metafeature-based Model Selection

Processing non-stationary data streams has gained a significant research interest in the recent years. Most of the current state-of-the-art approaches in the literature exhibit hybrid solutions based on an ensemble learning paradigm, extending its parallel architecture with a drift detector. Such integrated mechanism of concept drift detection often relies on the metadescription of incoming data. The article presents a proposition of a novel hybrid classifier – the Metafeature Concept Selector. The method employs the analysis of statistical metafeatures calculated on subsequent data batches to identify the currently processed concept. The work is concluded with an extensive experimental analysis of synthetic data streams with recurrent concepts, various dimensionality and various concept change frequencies, showing that the proposed approach brings improvement when integrated with state-of-the-art base incremental learners.