Community detection in directed networks based on network embeddings

Abstract

In real-world scenarios, many systems can be represented using directed networks. Community detection is a foundational task in the study of complex networks, providing a method for researching and understanding the topological structure, physical significance, and functional behavior of networks. By utilizing network embedding techniques, we can effectively convert network structure and additional information into node vector representations while preserving the original network structure and properties, solving the problem of insufficient network representations. Compared with undirected networks, directed networks are more complex. When conducting community detection on directed networks, the biggest challenge is how to combine the directional and asymmetric characteristics of edges. This article combines network embedding with community detection, utilizing the cosine similarity between node embedding vectors, and combining the ComDBNSQ algorithm to achieve non overlapping community partitioning of directed networks. To evaluate the effectiveness of the algorithm, we conduct experiments using both artificial and real data sets. The numerical results indicate that the algorithm outperforms the comparison algorithms (Girvan–Newman algorithm and Label Propagation algorithm) in terms of modularity, and can perform high-quality directed network community detection.