Research On Community Discovery Method Based On Autoencoder

Community structure is a prevalent network structure in everyday life,from simple networks to realistic and complex networks."Community" is a very important characteristic in complex networks.The use of community discovery technology is very important for studying the structural characteristics of complex networks,mining the implicit rules in real networks and predicting network behaviors,etc.Community structure often reflects certain implicit relationships and functions in networks,and its analysis can provide more information for deeper research and analysis.The work in this thesis focuses on the problems exposed by current community discovery algorithms such as inefficient model training,insufficient mining of node information and utilization of graph structure and clustering of multiple views,etc.Since traditional community discovery algorithms have many limitations and are not applicable on many large networks,we consider using deep learning methods for community discovery.Autoencoders are gaining importance in the task of community discovery because of their ability to efficiently represent nonlinear real networks.The provided embeddings are nonlinear in nature,so autoencoders are suitable for mapping high-dimensional data to low-dimensional data for community discovery requirements.Finally,we apply the community discovery algorithm to the home appliance industry chain service platform to cluster the types of services provided by the suppliers and the service attributes demanded by the demand side to recommend the desired services for them.The main work and research of this thesis is as follows:(1)Considering that the community discovery results obtained when using classical clustering algorithms to deal with high-dimensional matrices are not very satisfactory,we propose a community discovery algorithm based on an improved deep sparse autoencoder,and this algorithm is applied to the study of community discovery by combining two network similarity representations.Thus,the shortcoming that a single network similarity matrix cannot completely describe the similarity relationship between nodes is remedied.These similarity representations portray and consider as much as possible the local information among the nodes of the network topology.Then,weight-bound deep sparse autoencoders based on unsupervised deep learning methods are constructed and used to improve the efficiency of model training.Finally,feature extraction is performed using the similarity matrix to obtain a low-dimensional feature matrix and clustering using the k-means clustering algorithm to obtain plausible clustering results.Experiments demonstrate that this proposed method is more accurate than the communities discovered by directly using a single similarity matrix clustering algorithm,and the efficiency of the community discovery algorithm is significantly improved compared with other community discovery algorithms.(2)To explore the topology of data,many clustering methods embedded in graphs have been developed in recent years,but none of them consider the community-specific distribution of node representations,leading to unsatisfactory clustering performance.In addition,node attribute reconstruction and graph structure reconstruction are rarely considered simultaneously,resulting in reduced graph learning ability.We integrate node representation learning and clustering into a unified framework and propose a new graph attention auto-encoder for clustering,which learns more favorable node representations by using a self-attention mechanism and node attribute reconstruction,and at the same time,uses a l 1,2-parametric penalty to constrain the community-specific distribution of node representations so that nodes within the same cluster are represented as having the same distribution in the dimensional space,while nodes in different clusters are represented as having different distributions in the dimension.We modify the decoder reconstruction parameters to map the inner product of node pairs,which facilitates the decoder to reconstruct the graph structure better.Experimental results show that our proposed method outperforms several state-of-the-art methods in terms of performance.(3)In this thesis,we also do some research in multi-view clustering.In recent years,multi-view graph clustering has received deep attention in data mining,computer vision,pattern recognition,and has been widely used in recommender systems and face recognition.It seeks to partition the graph with multiple views to provide more comprehensive information.Although some research has been conducted in clustering multiple views and better results have been obtained,most of the current graph clustering algorithms are based on shallow models to deal with the information in the graph,which will severely limit the ability to model multiple views.We propose multi-view clustering on the graph attention autoencoder,which reconstructs multiple views using a single information view with content data,thus enabling the learning of node embeddings.In addition,a self-training clustering goal is proposed to improve the clustering results.Extensive experimental results show that our proposed method outperforms several state-of-the-art methods in terms of performance.