Identifying Structures with Informative Dimensions in Streams

Discovering structures in streaming data is an important data mining task and has motivated design of several well known algorithms. However, in some applications, a higher level of analysis is desirable to reveal the set of dimensions which contribute heavily to the structures. In this paper, we propose an algorithm ISID (Identifying Structures with Informative Dimensions), which operates in the streaming environment and delivers clusteres along with dimensions that contribute significantly to these clusters. The algorithm uses a three stage approach and utilizes entropy in an innovative way to achieve the goal in four different ways, depending on the desired guarantees on structural richness or minimal dimension set for a cluster.The experimental results on synthetic and real data sets demonstrate the efficiency and effectiveness of the proposed algorithm.