Research On Community Detection Method Based On Deep Learning

Community detection is an important task in complex network analysis.It helps to understand the characteristics of systems represented by complex networks and is of great significance for practical applications.Complex network structure is composed of plenty of nodes and perplexing relationships among those nodes,such as social network,collaboration network,biological network and transportation network.One of the main problems in complex network research is the detection of community structure,which has drawn the constant attention of researchers.Nowadays,many research achievements of community detection have been successfully applied to various fields,such as friend recommendation,personalized product promotion,protein function prediction,public opinion analysis and processing.Real-world networks are usually composed of functional units that are represented in the form of network modules or communities whose nodes are more closely connected than other network nodes.As an important feature of complex networks,community structure can reflect the important characteristics of the network to a certain extent.Therefore,identifying network communities is critical to describe organizational structures and understand complex systems.Community detection is dedicated to the accurate detection of community structures in complex networks.However,the network structure has become more complex due to the continuous expansion of network scale,and many traditional community detection methods have some defects when dealing with large-scale complex networks.In particular,when using classical clustering algorithms to process high-dimensional data matrices,the results are usually not accurate enough.To solve this problem,this thesis introduces deep learning into the research of community detection,aiming to extract effective low-dimensional feature representation of high-dimensional input data through deep learning framework,thus significantly improves the performance of community detection method.After an in-depth research on community detection and deep learning related theoretical knowledge,this thesis proposes two community detection methods based on deep learning:Community Detection Method based on Deep Autoencoder under the Influence of Transfer Learning(Transfer-CDDA)and Community Detection Method via Ensemble Clustering Framework(CDMEC).Through a large number of experimental on different datasets to prove the proposed methods in this thesis is feasible and effective,and it is compared with several existing community detection methods to further verify the good performance of methods.The main work of this thesis are as follows:1)Community Detection Method based on Deep Autoencoder under the Influence of Transfer Learning(Transfer-CDDA):Firstly,this method studies the similarity processing operation of network original data,that is,the transformation of adjacency matrix into similarity matrix.Inspired by the representation of the relationships between nodes,a new and effective network adjacency matrix transformation method is proposed to describe the similarity of nodes in the network topology.Then,a new community detection framework that named Community Detection Method based on Deep Autoencoder(CDDA)is proposed to extract nonlinear features,as so to obtain strong nonlinear characteristics of complex networks.Finally,in order to learn a more powerful representation of features,the unsupervised transfer learning is further introduced into CDDA(Transfer-CDDA)by minirmizing the divergence of Kullback-Leibler(KL)embedded instances,so as to ensure that the differences between different domains can be approximately equal when learning low-dimensional representation.At the same time,a new training strategy is proposed,in which the target domain and the source domain share the same parameters in the training process of deep autoencoder,and the training algorithm is optimized by using the back-propagation method of stochastic gradient descent.A large number of experimental results show that this method can guarantee good performance on both artificial benchmark networks and real-world networks.The method has more prominent advantages in the detection of complex community structures,2)Community Detection Method via Ensemble Clustering Framework(CDMEC):The method Transfer-CDDA adopts a single similarity matrix transformation method to describe the similarity relationship between nodes,and directly obtain the detection result by k-means clustering.Although this method is effective,the clustering results are not stable and there are defects in its applicability.In order to improve this defect and enable the method to fully describe the topology of complex networks,a method based on ensemble clustering framework(CDMEC)is proposed.Firstly,this method uses four different similarity matrices of complex networks are constructed by using four functions to fully describe the comprehensive similarity relationships between nodes in the network topology.Then,a mapping combination model of stacked autoencoder and transfer learning is established to obtain an effective low-dimensional feature information of complex networks.Finally,in order to improve the quality of the clustering algorithm,a ensemble clustering framework is proposed,which aggregates multiple inputs through the basic clustering algorithm to achieve the high-precision clustering performance of community detection.Plentiful experimental results show that the CDMEC algorithm is feasible and effective,which is similar to or even better than the optimal result obtained by the Transfer-CDDA method.