A Comparison of Techniques for Virtual Concept Drift Detection

Abstract

Concept Drift is one of the main problems presents in data stream processing for Data Mining and Machine Learning. This study focuses on Virtual Concept Drift. A common approach includes i) the detection of the drift with a specialized algorithm, and ii) the adaptation of the model to the current scenario. This work studies how well-known pre-processing methods affect abrupt Virtual Concept Drift detection in data streams. The proposed pre-processing techniques are: i) deleting the trend and ii) transforming the data stream from time to spectral domain. Moreover, three Virtual Concept Drift detection methods are compared over three publicly available data sets. According to the results, a slight improvement in the detection of Virtual Concept Drift is achieved when the trend is deleted. In contrast, no detection of Virtual Concept Drift is reported on the spectral domain.