Clustering Data Streams: A Complex Network Approach

Abstract

Clustering data streams is an interesting and challenging problem. Although several solutions have been proposed in the literature, some drawbacks remain. For instance, how to deal effectively with the offline process for partitioning the micro-clusters into macro-clusters is still an open problem. Typically, the k-means algorithm is considered in this phase, which despite precise results, require a mandatory user-defined parameter k, that defines the number of expected clusters. In this paper, we propose a new clustering method for data stream, named Prototype Networks. This method takes the complex network structure to represent the set of micro-clusters. This approach has proven to be advantageous mainly because these networks have an inherent community structure. As a consequence, the offline phase might be easily handled by a community detection algorithm, such as Infomap. The communities detected represents the cluster structure of the data assuming that the network construction was designed for this purpose. Computer experiments demonstrated the feasibility of the proposed approach. Moreover, the proposed method can detect automatically the number of clusters in evolving scenarios, which is a useful feature when dealing with data streams with concept drift.