Research On Network Community Detection And Representation Learning

The rapid development of information technology has brought us into the network era.Many objects in the real world can be represented as networks,such as social networks,power networks,and transportation networks.Community structure is one of the most important properties of complex networks,which has received widespread attention in recent years.A community usually represents a functional unit in a network.For example,in a protein interaction network,a community corresponds to a protein module with similar functions.Therefore,detecting the community structure in the network can help us understand the function of the network and mine the information hidden in the network.On the other hand,the traditional network representation generally uses sparse high-dimensional vectors,which greatly limit the use of machine learning methods,so that the calculation of high-dimensional vectors will consume a lot of computing resources when solving large-scale network problems.Network representation learning can represent nodes in the network as dense low-dimensional vectors,which can be effectively applied to tasks having important application value,such as visualization,community detection,and link prediction,which has important application value.In this thesis,a link clustering based memetic algorithm for overlapping community detection is proposed.Then,we conduct in-depth research on network representation learning methods.The main work is summarized as follows,The overlapping is indeed a significant feature of many real-world social networks,i.e.the same node may belong to different communities,which results in a sharp increase in the search space of node-based algorithms.However,link-based communities can more naturally represent overlapping structures because an edge can only belong to one community,which greatly simplifies the difficulty of coding.After the edge clustering is complete,the nodes connected to the edge naturally belong to different communities.We propose a new link-based community detection algorithm,named as Meme-Link,which uses the memetic algorithm to discover overlapping communities in the network.The experimental results on general and sparse networks show that our method can successfully detect overlapping community structures and almost all the overlapping nodes.Most of existing network embedding methods intend to preserve the pairwise relationship or similarity between nodes,but the community structure,which is one of the most important features of complex networks,is largely ignored.We propose a novel network embedding method based on evolutionary algorithm,termed as EA-NE_Community,which can preserve both the local proximity of nodes and the community structure of the network by optimizing a carefully designed objective function.The number of communities in the network can be automatically determined without any prior knowledge.Moreover,taking the intrinsic properties of network embedding problems in mind,a local search operator based on multi-directional search is designed which can effectively find feasible solutions.In the experiments,we first visualize the embedding representation obtained by different algorithms,and then use the problems of node clustering,node classification and link prediction to further validate the quality of the embedding representation obtained.Most of existing methods learn network representations only based on the structural features,but the rich attribute information of nodes is largely ignored.We propose a deep attributed network embedding method termed as DANE_Community,which is capable of learning node representations based on both structural and attribute information of the network.A novel adaptive weight balance mechanism is designed to combine these two kinds of information so that the weight between structural and attribute information can be automatically determined during the optimization process.Specifically,we take the community structure as a high-order proximity of nodes since it can solve the data sparsity issue.