Matching Degree Histogram-Based Concept Drift Detector for Data Stream

Data stream is becoming more common in various applications, and concept drift detection is an important work in data stream mining. Data distribution-based detection meth-ods are direct and effective. Specially, describing the distribution by histograms is unsupervised and applicable to various data. However, histogram is a relatively rough description of data, making it difficult to completely cover the information contained in data. Therefore, a matching degree-based histo-gram (MD-Histogram) is constructed firstly in this paper, in which the matching degree of each feature to each breakpoint is calculated. Besides, combined with the fix-slide windows model, a MD-Histogram-based concept drift detector (MDH-CDD) is proposed. MD-Histogram is used to describe the distributions of the data in different windows, and Euclidean distance is used to measure the difference. Then, according to different thresholds, three different states can be reported. The effectiveness of MDH-CDD is verified on synthetic and real-world datasets.