CUSUM Based Concept Drift Detector for Data Stream Clustering
K. Namitha, G. Santhosh Kumar · 2020
The last few decades mark an unprecedented growth in the number of applications producing high-speed data streams. Learning from such fast data streams has many inherent challenges. The dynamic change in the concept of the stream is a significant challenge to be handled by the learning systems. This problem termed concept drift is given due focus in data stream classification scenarios. But, data stream clustering algorithms usually treat concept drift implicitly as part of the learning process. The need for explicit drift detection and adaptation is often neglected. This paper discusses a statistical method of change detection for data stream clustering problems. The change detection is done based on CUSUM test. On identifying a change, the model is re-built using the recent samples from the stream. The change detection process has been validated using real and synthetic datasets.