Community Detection And Outlier Detection Based On Variational Autoencoder

Community detection is a valuable tool for understanding and analyzing complex social structures.Through Community detection,the hidden community structure in the original network can be revealed,so as to analyze the potential characteristics of the complex network.Community found that the mainstream of the algorithm is based on the network topology structure to explore the community,but in the real social network such as Facebook,due to abnormal connections such as spam or fishing accounts are visible,false node properties and the topological structure affect the community division,if not timely detection and even expand the rumor or the spread of the virus.At the same time,with the explosive growth of data information in recent years,feature dimensionality reduction in community has also become an important research topic in the network,which has attracted extensive attention of many scholars in recent years.However,most of the algorithms developed based on community dimensionality reduction make use of the classical theory,which takes a lot of time but does not work well for complex networks.Therefore,to correctly understand the community structure,it is very important to find the abnormal points that influence community clustering.At the same time,the application of efficient data dimension reduction method is of great significance to improve the quality of community partition.Therefore,this paper proposes a community detection and outlier detection method based on variational automatic encoder.First,an unsupervised outlier detection method based on graph embedding is proposed to effectively reduce the contribution of outliers to the total loss function based on the community connection structure and property characteristics,so as to optimize the graph embedding total loss function.Secondly,the core structure of the community can be found based on ktruss,and the range of k value can be initially limited by searching the core structure of different data sets.On the one hand,it ensures the core structure of the community,on the other hand,it speeds up the subsequent k-means and k-medoids clustering search for K value.Third,in order to meet the surging number of large networks,overcome the problem of"dimension disaster",the application of variational dimension reduction,automatic encoder to minimize the reconstruction error and KL divergence losses,finding the optimal solution at the same time,the effective use of local information and community training deep learning modules,for each vertex in the network embedded said.Fourth,based on k-means and k-medoids,the obtained low-dimensional data clustering can be used to obtain accurate community classificationCompared with the existing 5 algorithms of the same type,Fsame,NMI and modularity Q were used to analyze the quality of community partition.The experimental results obtained from Strike,Football,LiveJournal and Orkut showed good advantages on the four data sets with different data volumes.