Node Classification In Social Network Based On Dynamic Graph Encoder Network

The vigorous development of deep learning has caused a huge wave and Reform in the field of machine learning,such as intelligent voice answering,computer vision,data mining and information retrieval,and the classification of social network nodes is one of the branches.The data in these tasks are usually represented in Euclidean space.But now more and more data generated by applications can naturally evolve into graph,a novel data form,which is expressed in non Euclidean space.This is a great challenge to the traditional algorithms in related fields.However,in recent years,there are still many shortcomings in graph node classification methods:(1)these methods are only applicable to isomorphic graphs composed of the same type of nodes and edges,while the research on heterogeneous graphs with different types of nodes and edges is still insufficient.In this way,the social network diversified,complex data processing and analysis caused great obstacles,also unable to dig deeper the value of user information.(2)Traditional methods focus on mining local node features and graph structure information,but ignore the global continuity of node time series.Social networks are often dynamic.Most of the existing algorithms are based on static graph data structure,which can not effectively consider the integration of attribute information and interaction information between entities in different time steps of social networks.(3)These entities are not connected with all other entities.Only those entities that meet in the same field or for different reasons have such characteristics.Therefore,different nodes in social networks are connected and relatively independent.Aiming at the problem that it is difficult to consider the dynamic characteristics of social users in the existing static graph neural network methods,a social network node classification method based on heterogeneous dynamic graph model is proposed by introducing dynamic graph model.On the basis of dynamic graph modeling,the node feature updating mechanism based on point edge interaction and the time series aggregation method based on recurrent neural network are used to achieve efficient node classification of dynamic social network.The main work of this paper includes:(1)For some complex social relationships,isomorphic graphs usually can't fully represent the relationships among nodes.In this case,it is necessary to construct a heterogeneous graph with multiple types of nodes and their associated attributes.Unlike most of the existing work based on isomorphic graph,this paper proposes a neural network based on heterogeneous dynamic graph,which uses static graph encoder to process the node representation of each heterogeneous graph in different time steps,captures the changing characteristic information of nodes in all time steps of dynamic graph,and obtains effective node representation in each time step.After that,the recurrent neural network is used to mine the relationship and interdependence of different types of nodes in different time steps,so that the model can effectively learn to distinguish the characteristics of different types of nodes.Finally,it is verified on three kinds of real social network data,and the effect is better than the benchmark method based on static graph and dynamic graph.(2)Although social network is in dynamic development and continuous evolution,it has its own rules and characteristics,which means that there are certain connections between nodes in the same time step and different time step.On the basis of heterogeneous dynamic graph neural network,this paper proposes a sparse neural network based on heterogeneous dynamic graph to consider the connection relationship between entity nodes.The relationship between nodes is mined and learned through the sparse function,which is no longer the information aggregation of all nodes in different time steps in the traditional method.Finally,by comparing with the traditional static graph model,dynamic graph model and other sparse models on three kinds of social network data,it is proved that the proposed method is effective and feasible.