Autoencoder Based Community Detection with Adaptive Integration of Network Topology and Node Contents

Abstract

Community detection plays an important role in understanding the structure and laws of social networks. Many community detection approaches have been proposed and focus on topological structure alone. In addition to topology, node contents exist in real-world networks, and may help for community detection. Recently, some studies try to combine topological structure and node contents. However, it is difficult to address an inherent situation in real- world networks, that is the mismatch between topological structure and node contents in term of community patterns. When considering both topology and content of networks, the performance of those community detection methods is often limited by this mismatch. Besides, networks are often full of nonlinear features, making those methods less effective in practice. In this paper, we present an adaptive method for community detection, which is reached by a graph regularized autoencoder approach. This new method introduces a novel adaptive parameter to achieve robust integration of the topological and content information when there exists the mismatch between those two types of information in term of communities. Experiments on both synthetic networks and real-world networks further indicate that the proposed new method exhibits more robust behavior and outperforms the leading methods when there exists the mismatch between topology and content.