Processing data stream with chunk-similarity model selection

Pawel Ksieniewicz

Abstract: The classification of data stream susceptible to the concept drift phenomenon has been a field of intensive research for many years. One of the dominant strategies of the proposed solutions is the application of classifier ensembles with the member classifiers validated on their actual prediction quality. This paper is a proposal of a new ensemble method – Covariance-signature Concept Selector – which, like state-of-the-art solutions, uses both the model accumulation paradigm and the detection of changes in the data posterior probability, but in the integrated procedure. However, instead of ensemble fusion, it performs a static classifier selection, where model similarity assessment to the currently processed data chunk serves as a concept selector. The proposed method was subjected to a series of computer experiments assessing its temporal complexity and efficiency in classifying streams with synthetic and real concepts. The conducted experimental analysis allows concluding the advantage of this proposal over state-of-the-art methods in the identified pool of problems and high potential in practical applications.