Hypergraph Embedding Overlapping Community Discovery Based On Comprehensive Community Walk

In the field of complex network research,for different application scenarios and increasingly complex network topologies,it is an important research topic to study a universal community discovery method.In the traditional community discovery method,the rule of community division is usually designed according to different application requirements.Its effectiveness and universality are difficult to meet the needs of current largescale complex network analysis applications.The network representation through the embedding of network topologies is a new approach to address community discovery.However,how to represent the community features with the network structure,especially the overlapping community features,is the key to effectively conduct general community discovery.This paper proposes an embedding method based on comprehensive community walk to learn community structure features,and to conduct the overlapping community discovery by the hypergraph transformation.A community discovery method based on comprehensive community embedding is proposed in this paper,which provides a new idea for feature learning methods adapted to solve different downstream tasks.In order to capture the community features of the adjacency,local modularity and edge betweenness of the network structure,the proposed method is based on the community similarity assumption and embeds the community features in the community walk process.Combined with the community structural features and node attributes,a comprehensive community representation of nodes is realized.Then,the K-clustering neural network model is proposed to realize community division by means of unsupervised node clustering and to optimize the number of communities and the quality of community discovery in an iterative process.Experimental results on the datasets of social network and co-author network show that the proposed method performs stably and outperforms traditional methods,especially,the evaluation of modularity on Coauthor A＿Net Att dataset reaches 0.757.Aiming at the representation of overlapping structure features,the overlapping community discovery method based on hypergraph embedding is put forward,which introduces the overlapping features to the previous comprehensive community walk.This method utilizes the hypergraph transformation to conduct a community walk,with hypergraph events as the object,and to learn the community feature vectors of events.Through the K-clustering neural network model,the community vector representation of events is clustered to obtain a non-overlapping community structure by using the event as community members.Finally,the overlapping community structure is obtained according to the affiliation of hypergraph events and network nodes.The experimental results on real network datasets of different scales show that the overlapping community discovery method is better than others in the evaluation of accuracy and overlapping quality.In particular,overlapping modularity and the best F1 score on Cora dataset respectively reach 0.610 and 0.880.There are 32 figures,14 tables,and 77 citations in this thesis.