Granularity Adaptive Density Estimation and on Demand Clustering of Concept-Drifting Data Streams

Clustering data streams has found a few important applications. While many previous studies focus on clustering objects arriving in a data stream, in this paper, we consider the novel problem of on demand clustering concept drifting data streams. In order to characterize concept drifting data streams, we propose an effective method to estimate densities of data streams. One unique feature of our new method is that its granularity of estimation is adaptive to the available computation resource, which is critical for processing data streams of unpredictable input rates. Moreover, we can apply any clustering method to on demand cluster data streams using their density estimations. A performance study on synthetic data sets is reported to verify our design, which clearly shows that our method obtains results comparable to CluStream [3] on clustering single stream, and much better results than COD [8] when clustering multiple streams.