Kernels on Attributed Networks for Community Detection

Community detection is a primary problem in the study of complex networks. When graphs are enriched with attributes, it has been found that this additional information can help in better understanding the ties among the actors composing the network and provides a deeper insight into group organization. The paper proposes the investigation of a multiobjective genetic algorithm for attributed networks extended with kernel functions for computing node similarity both in terms of structure and features. The commute-time kernel, based on the concept of random walk, is first applied to the adjacency matrix of the graph and then four kernels are applied for computing the similarity between nodes while simultaneously optimizing structure and feature dimensions. Simulations on both synthetic and real-world citation networks show that kernels effectively improve the quality of the resulting partitions.