Seeding the Kernels in graphs: toward multi-resolution community analysis

Current endeavors in community detection suffer from the resolution limit problem and can be quite expensive for large networks, especially those based on optimization schemes. We propose a conceptually different approach for multi-resolution community detection, by introducing the kernels from statistical literature into the graph, which mimic the node interaction that decays locally with the geodesic distance. The modular structure naturally arises as the patterns inherent in the interaction landscape, which can be easily identified by the hill climbing process. The range of node interaction, and henceforth the resolution of community detection, is controlled via tuning the kernel bandwidth in a systematic way. Our approach is computationally efficient and its effectiveness is demonstrated using both synthetic and real networks with multiscale structures.