Research Of Multi-layer Network Community Detection

As the development of technology and informatization progresses,the multidimensional nature of the connections between entities in the real world is more obvious.For example,in social networks,different social platforms own diverse socialization properties since they represent various social preferences.Therefore,how to represent the social behavior of users under different social platforms is meaningful for social network analysis.Due to the unique hierarchical structure,multi-layer networks can represent the multi-dimensional connections between entities accurately.Therefore,multi-layer networks become an important tool for modeling,analyzing and studying multi-relational complex systems.In network science,identifying community structures of multi-layer networks is important for exploring node functions and revealing the potential structures of networks.For instance,detecting the community structure of social networks by combining the information of different social platforms can help researchers explore the potential social circles,which is conducive to addressing downstream tasks such as hot topic prediction and rumor prevention.In view of the important practical significance of multi-layer network community detection,this paper focuses on the following two research contents.Firstly,guiding the multi-objective evolutionary algorithm based on the consensus prior information to ensure the accurate community detection in different network structures.Secondly,extracting and aggregating the information of multi-layer networks to achieve accurate community detection without the guidance of prior information.(1)There are some issues that existing algorithms depend only on the topological structure,resulting in the low accuracy of networks with complex structures.To address the problem mentioned above,this paper proposes a multi-objective evolutionary algorithm based on consensus prior information.The proposed algorithm applies the consensus prior information to guide the evolutionary computation process from different aspects,so that it can achieve high-precision community detection.Firstly,the algorithm detects the graph-level information based on Node2 vec and density-based aggregation strategy,as well as node-level information based on Jaccard similarity,respectively.Furthermore,the proposed algorithm generates the prior layer and initial population according to graph-level information,which can guide the algorithm from a global perspective.The node-level information is applied to guide the mutation strategy to help generate the optimal population.Moreover,after the in-depth analysis of multi-layer network structures,the modularity obtained by the fast clustering method is applied to quantize the community structure clearness of each layer,so as to make full use of the valid information and reduce the influence of noisy information.Finally,the accuracy and robustness of the proposed algorithm are verified based on different kinds of datasets.(2)Due to the computational cost of the consensus prior information extraction process affecting the scalability under large-scale networks,this paper proposes a network augmentation contrastive constraint method for multi-layer network community detection.The proposed algorithm achieves accurate community detection by contrasting the augmented network and multi-layer network.Firstly,the proposed algorithm constructs a network generation model and a node representation model.Based on the network generation model,the information of the multi-layer network is fused to generate an optimizable augmented network.Then the algorithm computes the low-dimensional representations of the augmented network and each layer of the multi-layer network based on the node representation model.Finally,the algorithm updates both models by contrasting the representation between the augmented network and multi-layer network.Therefore,the algorithm can improve the robustness of the node representation model and the feature aggregating ability of the node representation model.Experiments show that the proposed algorithm performs well on various datasets.