Research On Network Embedding Method Based On Graph Attention Mechanism

The rapid development of the big data era has led to an exponential and rapid increase in the volume of data from all types of complex networks,making the attributes of network data more diverse and the structure more complex,which makes the data difficult to process.In order to analyze and process data efficiently,the research of network embedding is crucial.Network embedding,also known as graph embedding,abstracts network data into a graph structure,using nodes in the graph to represent network data points and edges in the graph to represent relationships between data points.The feature information of the graph is extracted and embedded into a low-dimensional vector space to obtain a dense representation of the network,which can later be used for applications such as community discovery,node classification and link prediction of the network.Existing network embedding methods are singular in their consideration of the attributes and structure of the data.To address this issue,this paper proposes two network embedding methods that fuse multiple aspects of the data and use the resulting network representations for community discovery and node classification respectively to verify the efficiency of the representations.The main research of this paper is as follows:(1)Network embedding method based on label attention mechanism is proposed.Firstly,the network nodes are divided into label nodes and example nodes based on the idea of label distribution.Then,designing a new representation reflecting the global and local neighborhood information of the network nodes,transforming the features obtained through label frequency and inverse example frequency processing into a feature matrix,introducing an attention mechanism to extract the feature information of the matrix in order to better assign weights to the node neighbors,and composing it into an attention feature matrix so that the network acquires more detailed information related to the target to achieve a more accurate network embedding.The feature matrix obtained from the embedding was subsumed after network community pre-processing and community gaming to do the community discovery task.The number of community divisions,normalized mutual information,modularity and running time are compared on seven real network datasets and the results of the divisions are visualized and analyzed,and the results show that the proposed embedding method has better performance.(2)Capsule network embedding method based on neighborhood attention mechanism is proposed.Firstly,discovering network structure information and fusing node aggregation coefficients and node connection strengths using the network adjacency matrix to discover neighborhood similarities in network topology information.Secondly,the network feature information is mined,and the network feature matrix is processed using the attention mechanism on the neighboring node features so that the network features contain more information between attributes;again,the network structure and feature information is fused,and the fused information is processed by convolution to generate node capsules,which are used as input to the capsule network,and the capsule network representation of the nodes is generated through the transmission of the capsule network.Finally,a minimisation Softmax loss function is trained on the model to learn the final network representation,which is fed into a regularized logistic regression classifier to classify the network nodes.Citation networks are attribute networks and data nodes have class labels.A comparison of the classification accuracy of nodes with different numbers of routing iterations and connection strengths for different neighborhood orders on three datasets of citation networks shows that the proposed embedding method improves the classification accuracy of the node classification task.