Engineering graph clustering: Models and experimental evaluation

Reviewer: Goran Trajkovski

Clustering is an issue that affects nearly all computing disciplines, from computer graphics to bioinformatics; it also affects a wide range of fields outside of computing. Trying to identify regions in which data are denser than other regions is not a trivial task when the dataset is huge. So what do we do then__?__ We try to formalize intuitive ways to see the denser regions in the foggy clouds of data. Effective, well-performing algorithms are a necessity for data mining. In this paper, the authors give us models of graph clustering, based on the notion of identifying intracluster density as well as intercluster sparsity. They contrast three algorithms: Markov clustering, interactive conductance cutting, and the geometric minimum spanning tree clustering. The authors define several parameters for quality control, in an effort to assess density and sparsity. For graphs of different topologies, one algorithm might perform better than the other two. As always, there is a tradeoff between complexity and speed. Experimental studies reveal that this approach is a promising one. So what is the effect of this paper__?__ It makes one sit down and think. Given time, one might implement an approach, and apply it to datasets that have been collected and hidden somewhere, in order to see how the algorithms discussed in this paper will perform. This is a well-written, organized, and easy-to-follow paper. It presents new concepts with sound fundamental investigations and experimental support. It represents a vast, solid amount of work. Online Computing Reviews Service