Community Detection in Dynamic Networks: A Novel Deep Learning Method

Dynamic community detection has become a hot spot of researches, which helps detect the revolving relationships of complex systems. In view of the great value of dynamic community detection, various kinds of dynamic algorithms come into being. Deep learning-based algorithms, as one of the most popular methods, transfer the core ideas of feature representation to dynamic community detection in order to improve the accuracy of dynamic community detection. However, when committing feature aggregation strategies, most of methods focus on the attribute features but omit the structural information of networks, which lowers the accuracy of dynamic community detection. Also, the differences of learned features between adjacent time steps may be large, which does not correspond with the real world. In this paper, we utilize the node relevancy to measure the varying importance of different nodes, which reflects the structural information of networks. Having acquired the node representations at each time step, the cross entropy is used to smoothen adjacent time steps so that the differences between adjacent time steps can be small. Some extensive experiments on both the real-world datasets and synthetic datasets show that our algorithm is more superior than other algorithms.