Integrating heterogeneous structures and community semantics for unsupervised community detection in heterogeneous networks

Community detection aims to discover hidden communities or groups in complex networks and is essentially unsupervised clustering behavior. However, most of the existing unsupervised methods are designed for homogeneous networks; therefore, they cannot effectively handle heterogeneous structures and rich semantic information. Under such a situation, it is difficult to accurately detect communities in heterogeneous networks that better reflect the real world. Therefore, this work aims to design an unsupervised framework to fuse heterogeneous structure information and interpret the rich semantics of the network in the form of community semantics. Thus, a heterogeneous network community detection method, called HAESF, is introduced. It includes two modules: the Heterogeneous Auto Encoder (HAE) and the Semantic Factorization (SF) modules. In more detail, the HAE module adopts a hierarchical attention scheme to represent and aggregate the heterogeneous structure of the network. And it proposes the concept of heterogeneous information combinatorial graphs for structural reconstruction to achieve unsupervised detection. Concerning the SF module, it focuses on learning the semantic information in the network from the community point of view. It uses nonnegative matrix factorization to decompose the network features for obtaining community semantics. Once both modules are implemented, the objective of restricting community segmentation based on these semantics is achieved. The constraint is based on community semantic homogeneity to correct inaccurate node delineation. Furthermore, to improve the algorithm efficiency, a unified framework is designed to optimize the HAE and SF modules jointly. Within this new framework, the SF loss is innovatively used as a judgmental loss for selective segmentation optimizations, helping to obtain more reliable community detection results. As for the results, extensive experiments are performed on three public datasets. The findings show that HAESF outperforms the other popular unsupervised methods, where the composite score of HAESF is 11.73% ahead of the next best, demonstrating the proposed method’s effectiveness.