A Density-Based Clustering Approach for Behavior Change Detection in Data Streams

Mining data streams poses many challenges to existing Machine Learning algorithms. Algorithms designed to learn in this scenario need to constantly update their decision models in accordance with current data behavior. Therefore, the ability to detect when the behavior of the stream is changing is an important feature of any learning technique approaching data streams. This work is concerned with unsupervised behavior change detection. It suggests the use of density-based clustering and an entropy measurement for change detection that is independent of the number and format of clusters. The proposed approach uses a modified version of the Den Stream algorithm that is designed to better cope with the entropy calculation. Experimental results using synthetic data provide insight on how clustering and novelty detection algorithms can be used for change detection in data streams.