Research On Heterogeneous Network Representation Learning Based On Graph Attention Mechanism

Network representation learning is a study that explores how to better represent network information.It represents the nodes in the information network as lowdimensional dense real vectors for downstream machine learning tasks.In recent years,network representation learning has attracted the attention of a large number of researchers due to its wide range of application scenarios,and network representation learning that incorporates heterogeneous information has also become an important research hotspot.The initial heterogeneous network representation learning mostly uses meta-path and random walk methods,which can extract the rich semantic information in the heterogeneous network while retaining the local structure information of the network,However,these methods usually ignore the different importance of the different neighbors of the nodes.In addition,meta-paths often need to be designed by experts,and in the process of using meta-path based sampling,intermediate nodes may be skipped due to the constraints of meta-paths,resulting in the loss of information on the way.In recent years,graph attention networks have been proposed to consider the importance of different nodes,and solve the problem of feature weighted aggregation in graph structure.However,it is only suitable for homogeneous networks,and it is difficult to retain structural information in the network.On the other hand,most of the methods based on graph neural networks are semi-supervised,it is difficult to make full use of the information contained in unlabeled data.In order to solve the above problems,this paper explores the representation learning of heterogeneous information networks based on graph attention mechanism.The main work proposed in this paper is:1.Aiming at the problem that the meta-path-based method ignores the different importance of nodes,and the pure graph attention network considers the weight of the nodes,but it is only suitable for homogeneous networks and it is difficult to retain the structural information of the network.This article is in On the basis of using meta-path to extract semantic information of heterogeneous networks,the feature of aggregation nodes in graph attention layer is added,and the information sampled by meta-path is merged into graph attention network,so that node information can be better aggregated.Experimental results on classic data sets of heterogeneous networks such as DBLP show the effectiveness of the proposed method.2.Aiming at the problem that meta-paths usually require expert design and it is easy to lose part of the information passing by,this paper proposes a heterogeneous information network representation learning model that does not use meta-paths,and directly applies graph attention networks to heterogeneous networks.Represents learning.Regarding the problem that the graph attention network is difficult to retain the network structure information,this paper uses the node representation after training of the random walk sequence to calculate the similarity between nodes,and sends it to the graph attention as an additional structural similarity coefficient.Participate in learning online.Experimental results on classic data sets of heterogeneous networks such as DBLP show the effectiveness of the proposed method.3.Aiming at the problem that the semi-supervised task based on graph neural network cannot make full use of a large amount of unlabeled data,this article draws on the multi-task learning method in supervised learning,and adds different auxiliary tasks on the basis of heterogeneous graph attention network.The structure and attribute information of the network itself is transformed into supervision information to constrain the training of the model,so as to learn more complete information in the network through this use of unlabeled data.Experimental results on classic data sets of heterogeneous networks such as DBLP show the effectiveness of the proposed method.