Research On Dynamic Network Representation Based On Community Information

Network representation is a preprocessing link in analytical tasks of complex network,which maps entities to low-dimensional feature space and preserves the original information in the network.The existing methods of network representation mostly preserve the local structure information in the network,and do not consider the role of global information,such as community information.In addition,the community of multi-granularity is of great significance to network representation.To this end,the first research goal of this thesis is to preserve more multi-granularity information in the static network representation.Practically,the network is often dynamic and its evolution enormously affects the community,the change of the community information further affects the representation results of the original network.Thus,fusing community information in dynamic network representation is the other aim of this thesis.The main tasks of this thesis are as follows:1.A network representation method based on multi-granularity community information is proposed.This method can both preserve the structure information and learn the network’s multi-granularity community information.Based on the Skip-gram model,the embedding method can capture the local structure information of the network;combine the community detection algorithm to obtain the multi-granularity community structure of the network,and calculate community embedding for the community in each granularity and the weight between the communities;according to the community embedding,the joint optimization method is used to adjust node embedding,and preserve community information in each granularity;finally,splicing is performed to obtain a representation result that fusion of multi-granularity community information.To verify the effectiveness of the method,link prediction and node classification experiments were conducted on four real networks.The experiment proved that this method can more effectively improve the accuracy of downstream tasks compared with other representation methods that only consider structural information or single-granularity community information.2.A dynamic network representation method based on community information is proposed.This method can adjust the representation of the network at the last moment according to the evolution of the community in the context of a dynamic network and fuse the community information of the current network.This method uses the dynamic network representation method to learn embedding of incremental nodes,and preserves the structural information of the network;it uses the dynamic community detection algorithm to obtain changed communities and captures the community information of the current network;it calculates the community embedding results,and uses the joint optimization method to update the embedding of the node where the community attributes changes so that the embedding results can be fused with the community information of the current network.This method conducts link prediction experiments on four real dynamic networks and proves that additional consideration of community information in dynamic network representation can improve the accuracy of predicting future links.3.For the two methods,a prototype system that can run two network representation methods is designed and implemented.The system can use the corresponding representation method according to the parameters set by the user to obtain the network representation result and display it.