Concept Drift Detection via Equal Intensity k-Means Space Partitioning

Data stream poses additional challenges to statistical classification tasks because distributions of the training and target samples may differ as time passes. Such distribution change in streaming data is called concept drift. Numerous histogram-based detection methods have been proposed detect Most histograms are developed on grid-based or tree-based space partitioning algorithms which makes partitions arbitrary, unexplainable, cause drift blind-spots. There a need improve accuracy for with unsupervised setting. To address this problem, we propose cluster-based histogram, equal intensity k-means (EI-kMeans). In addition, heuristic method sensitivity introduced. The fundamental idea improving minimize risk creating offset regions. Pearson's chi-square test used hypothesis so that statistics remain independent sample distribution. number bins their shapes, strongly influence ability drift, determined dynamically from based an asymptotic constraint test. Accordingly, three implement detection, including greedy centroids initialization algorithm, cluster amplify-shrink algorithm. For adaptation, recommend retraining learner if detected. results experiments synthetic real-world datasets demonstrate advantages EI-kMeans show its efficacy detecting