Research On Community Discovery Methods For Different Kinds Of Networks

Community detection is one of the critical technology in the network analysis area.By leveraging rich information in networks,community detection can effectively and accurately uncover the node clusters(a.k.a communities)within the network,achieve the goal of profoundly understanding the essential relationship in the target network.At the same time,with the more in-depth exploration of deep learning related technologies in network analysis area,community detection has become one of the vital analysis topics in the network analysis area.As is known to all,to form a community,the nodes within a community should share not only tight connections but also have similar attributes.However,most of the current community detection methods only focus on connections and weights among node pairs,while ignoring other relevant information,such as nodes attribute,node pair relationships in multilayer networks,e.t.c.As a result,the uncovered communities can only reflect the tightness between node pairs on topological relationships.To fully use given information within a network,and to increase the accuracy of a community detection method,this thesis targets at four different types of networks(timedependent networks,multilayer network,network with node attributes,and network with node attributes and labels).By combining advantages in deep learning algorithms such as network embedding or graph neural network,this thesis proposes a suitable community detection method for each type of network.The main content and innovation of the paper mainly summarized in the following five aspects:1.Community detection in time-dependent networksAs a type of dynamic network,in time-dependent networks,the topology changes between two timestamps are always constraints in a relatively small area of the network.Base on such understanding,this thesis proposes an incremental community detection method to discover communities for each timestamp in time-dependent networks.By considering communities detected from the previous timestamp,this method can efficiently and accurately discover communities for the current timestamp.2.Community detection in multilayer networksMultilayer networks have been introduced to capture each type of relationship among a group of nodes,where each layer indicates a type of relationship.To co-analyzing communities from multilayer networks,this thesis proposes a community detection method for multilayer networks.By capturing the topological relationship among nodes in the multilayer networks,the method constructs an edge-based hierarchical structure for the multilayer network.Then,by proposing community density in a multilayer network,this method can measure the overlapping community structures for the target multilayer networks.3.Network embedding for multilayer networksThis thesis proposes a new network embedding technology,which extends the current network embedding methods initially suitable for analyzing a single network to multilayer networks.Specifically,by introducing a new hyper-parameter to guide the random walk process to transfer between layers in multilayer networks,this method can capture the topological relationship not only in a single layer,but also multiple layers.4.Nodes attribute-aware community detectionThe real world network has not only to contain topological information,but also have attributed information for nodes.This thesis proposes a network embedding method which can perceive node attributes.Firstly,according to the node's attribute information,this method constructs the node's neighbor relationship in the attribute space.Secondly,by using random walks to analyzing the topological structure of nodes,this method forms the node's neighbor relationship in the topology space.Finally,through a unified optimization function,this method project nodes into a unified vector space which can describe the joint relationship between nodes in attribute space and topological space.In addition,to adapt to time-dependent networks,this paper also proposes an incremental process to adopt the proposed method for time-dependent networks.5.Nodes label-aware community detectionGiven a network with partial labels,how to detect other labels has become one of a hot topic in semi-supervised community detection methods.By using graph neural network,this thesis proposes a novel community detection method which can leverage the partial information on given node labels.Specifically,by using trainable non-negative matrix decomposition method to analyze the hierarchical structure of the input network topology,the proposed method uses high-order neighbor information of the targeted network as an expanded receptive field of a graph neural network.Then,by applying an optimization function,the proposed method is able to predict labels information in an end-to-end training manner.In general,this method predicts the unknown labels in the network based on a small number of given labels,to obtain the community structure of the target network.The empirical results show that the methods proposed in this thesis can effectively and accurately identify the community structure for each kind of network.