Research On Dynamic Network Representation Learning Methods

Network representation learning aims to learn a mapping function,which can map all nodes in the network to low-dimensional vector representations,so as to extract the characteristics of nodes and carry out network analysis.In recent years,more and more network representation learning methods have emerged,and most of the existing methods only study static networks.However,in real networks,the structure and attributes of networks are always changing over time.Therefore,it is a promising research subject to make use of the time-varying network topology structure and attributes to mine the characteristics of complex network,and to learn the dynamic node embeddings which can describe the change rule of network and be able to utilized to conduct complex network analysis tasks.There are two main challenges in the research of dynamic network representation learning.On the one hand,there may be anomalous edges in the dynamic network,if network representation learning is based on the information which include anomalous edges,then the learned node embeddings cannot depict precise network properties effectively.In addition,anomaly labels are scarce,and it is impractical to realize supervised learning based on anomaly labels,therefore,it is challenging to learn robust node embeddings that can be used to detect network anomalies in the case of scarce labels.On the other hand,most researches on dynamic networks are carried out by dividing them into multiple static network graphs,which ignores the continuously changing timestamp information in the network and leads to the loss of information in the dynamic network.In view of the aforementioned problems,this paper proposes two works.Firstly,a dynamic network representation learning method for anomalous edge detection(Dynamic Network Embedding for Anomaly Detection,DNEA)based on deep learning is proposed.Complex networks generally possess community structure,and since the anomalies in the network are more likely to deviate from the community division of the real network,it is necessary to analyze the community structure of the dynamic network through the observed topology structure.DNEA captures the community structure of dynamic network by studying the relations between nodes and neighbors,and reconstructs the network according to the results of block partition.In this study,the reconstruction ability of an edge is taken as its anomaly score to verify whether an edge is anomalous or not.In order to solve the problem of scarce labels,the model uses the method of negative sampling,which can be trained end-to-end to learn the robust representation of nodes in the dynamic network.Finally,the feasibility of the model is verified by experiments on three real dynamic network datasets.Then,a novel dynamic attributed network representation learning method(Node Embedding over Dynamic Attributed Network,Dyn ANE)is proposed.In this study,the structural information and attribute information of dynamic network are combined to capture the continuous-time and discrete-time changes of the network respectively.When mining continuous-time changes,the model takes into account the timestamp information of node interaction,and believes that the more recent the node interaction,the greater the influence on the node representations.When mining discrete time changes,the model captures the community structure through the neural network.The neighbor structure and community structure mined by the two kinds of changes are integrated into the node representation so that the learned representation contains the rules of network dynamic change.Finally,the model is compared with other algorithms through the link prediction task,which proves the superiority of the proposed model.