Energy-Based Clustering of Graphs with Nonuniform Degrees

Widely varying node degrees occur in software dependency graphs, hyperlink structures, social networks, and many other real-world graphs. Finding dense subgraphs in such graphs is of great practical interest, as these clusters may correspond to cohesive software modules, semantically related documents, and groups of friends or collaborators. Many existing clustering criteria and energy models are biased towards clustering together nodes with high degrees. In this paper, we introduce a clustering criterion based on normalizing cuts with edge numbers (instead of node numbers), and a corresponding energy model based on edge repulsion (instead of node repulsion) that reveal clusters without this bias.