Graph embedding based on motif-aware feature propagation for community detection

Identifying communities plays an essential role in disclosing modular structures with specific functions or properties in different kinds of complex networks. However, the low accuracy can be attributed to the fact that traditional algorithms always pay more attention on networks’ structural information of lower-order, i.e. neglect that of higher-order. In this paper, we propose a novel algorithm, Graph Embedding based on Motif-aware Feature Propagation (GEMFP) for community detection. In graph embedding, the vector representation of each node in a graph is firstly initialized at random. Then, we reconstruct a weighted graph instead of the original one by considering the information of lower-order (adjacency interactions) as well as higher-order (motifs), which is further used to define the influence acceptance between nodes, and the representation of a node is determined by aggregating the information of its neighbors’ influence iteratively until the convergence of this process is reached. Finally, K-means algorithm is adopted to extract communities by using the above embedding information. We conduct extensive experiments on 14 real-world datasets, and the results show that our algorithm tends to be more effective for community detection compared with several traditional, and graph embedding algorithms. The code for GEMFP is accessible on GitHub at https://github.com/zhanghan1020/GEMFP.