Node Representation Learning In Heterogeneous Information Networks Based On Generative Adversarial And Graph Attention

As the large-scale information network represented by social network emerges in an endless stream.How to make full use of this information to mine a common representation suitable for various tasks is particularly important.In fact,large-scale information networks often contain a lot of complex interaction and semantic information,and have multi-source heterogeneity.This brings a challenge to the traditional network representation learning methods.The emergence of deep learning has opened up a new path for network representation learning.It greatly promotes the development of network node representation study.Based on the graph attention mechanism and generative adversarial networks,this paper researches on heterogeneous information networks and hypergraph networks to solve the problem of insufficient node representation ability.The specific research content includes:(1)A heterogeneous network node representation learning algorithm(Adversarial Variational Graph Autoencoder,AGVAE)based on adversarial graph autoencoder is proposed.The global structure of the network is captured by using adversarial graph autoencoder and the low dimensional dense representation vector of nodes is generated.Among them,the encoder is implemented by a two-layer graph convolutional network,which aggregates the neighborhood features of nodes by means of convolution,and reconstructs the network as the input of the decoder.Secondly introduced adversarial training,generation model section and encoder share the same network.Discriminant models can be thought of as a binary classification task.It is mainly used to distinguish the false samples generated by the generation model from the true samples sampled from the real data set.According to the results,the parameters of the generated model are updated in reverse so as to optimize the model.Finally,in the joint optimization stage,the regularization constraint is added,and the penalty coefficient is used to increase the punishment of bad results,so as to achieve the goal of optimization.(2)Heterogeneous hypergraph network node representation learning algorithm(Hypergraph Attention Autoencoder,HATAE)which integrates external information is proposed.On the basis of the above research method,this method further excavates the multiple correlation relations in heterogeneous information networks.By focusing on the higher order topological structure information and the lower order nearest neighbor information,the multi-correlation relationship pattern in the heterogeneous information network is matched by the mode extraction,and then the subgraph of the heterogeneous information network is extracted and named as the heterogeneous hypergraph network.This paper presents a hypergraph attention mechanism based on autoencoder to uniformly encode node information and attribute information in heterogeneous hypergraph networks.The hyperedge correlation matrix is reconstructed by the decoder to preserve the global attribute information of the node.At the same time,the node neighbor function is added to judge whether different types of nodes belong to the same hyperedge,and the first-order neighbor is reserved.The learned node representation is applied to the hyperside link prediction of the downstream tasks,and the robustness of the representation results is evaluated by the actual mining tasks.In a homogeneous network dataset Cora and three original heterogeneous network GPS,Movie Lens and Drug datasets conduct experiment.For the hypergraph network node representation learning,the data set adopted is the extracted heterogenous hypergraph network subgraph.A comparative experiment with the classical representation learning method is completed.The results prove that the AGVAE and HATAE algorithms are superior to the traditional methods in link prediction.The effectiveness of the proposed algorithm is fully verified.