Dynamic community detection over evolving networks based on the optimized deep graph infomax

As complex systems, dynamic networks have obvious nonlinear features. Detecting communities in dynamic networks is of great importance for understanding the functions of networks and mining evolving relationships. Recently, some network embedding-based methods stand out by embedding the global network structure and properties into a low-dimensional representation for community detection. However, such kinds of methods can only be utilized at each single time step independently. As a consequence, the information of all time steps requires to be stored, which increases the computational cost. Besides this, the neighbors of target nodes are considered equally when aggregating nodes in networks, which omits the local structural feature of networks and influences the accuracy of node representation. To overcome such shortcomings, this paper proposes a novel optimized dynamic deep graph infomax (ODDGI) method for dynamic community detection. Since the recurrent neural network (RNN) can capture the dynamism of networks while avoiding storing all information of dynamic networks, our ODDGI utilizes RNN to update deep graph infomax parameters, and thus, there is no need to store the knowledge of nodes in full time span anymore. Moreover, the importance of nodes is considered using similarity aggregation strategy to improve the accuracy of node representation. The experimental results on both the real-world and synthetic networks prove that our method surpasses other state-of-the-art dynamic community detection algorithms in clustering accuracy and stability.