Research On Node Classification Problem Based On Graph Embedding

Social networks serving as an important source of data can be seen as a virtualized society that realizes the connection between people,things,and things.For example,the network constituted by the user mail communication,the network composed of the user selected commodity,and the network composed of mutual reference between the papers,etc.Currently,most of the research on social networks is based on graph embedding techniques.The so-called graph embedding,also known as graph characterization learning.It is essentially a process of compression of the network graph and a vector representation.Many downstream tasks are completed on this basis,such as node classification,clustering,connection prediction,etc.Social networks provide important data for scholars to study the node classification problem.This paper is based on the relevant study of node classification problems on social networks.In the study of the node classification problem,how to improve the correctness rate of the node classification is the goal of the long-term efforts of scholars.This paper investigates the problem of node classification for static networks.SNE.,a node classification algorithm based on latent information and characterization learning The proposed approach yields a new network topology through a second-order similarity among the nodes.This network topology enhances the association between nodess and improving me the effect of node classification in the network.However,most social networks in are dynamic.Such as increase or decrease of nodes,change in the relationship between nodes,etc.Because of the existence of these factors,the research method of static networks will no longer apply to dynamic networks,so this paper studies the node classification problem of dynamic networks from two aspects.First,AENC,a node characterization learning algorithm based on self-coding framework The proposed method considers embedding the network matrix with neighbor relations under the first time snapshot in a dynamic network into a low-latitude vector space.Then,through self-coding framework,obtain the vector representation of the last moment node,improve the learning effect of node characterization,and then improve the node classification effect.At the same time,the important is that the paper proposes a simple method to calculate the stability of the node characterization.The smaller the calculation results,that is,the more stable the node representation obtained by the algorithm,the more importance to improve the efficiency of the node classification.Secondly,ENGRU,a node classification algorithm based on GRU model and self-coding model Network relational information with a time snapshot is embedded into low-latitude vectors via RNN models and then decoded through a self-coded decoder.The coding process of this model can make better use of historical information to learn better node representations.And the reconstructed network obtained by the decoding process can better train the model,thus improving the node classification task.