Dual-Channel Attention Mechanism Semi-Supervised Graph Learning Network

Graph data is a common non-Euclidean data structure,which is used in every aspect of human life.In recent years,as a powerful deep learning method,graph neural network has shown remarkable ability in solving graph learning problems.Most graph neural networks usually obtain feature representation by equal-weight aggregation of neighbor information of nodes.However,the performance of traditional graph structures for node features and topological structures is not ideal,which makes graph neural networks unable to obtain more comprehensive graph information,and the models designed in this way usually have disadvantages such as poor flexibility.A lot of label information is often required for training graph neural networks,which is difficult to obtain in real life,which obviously limits the wide use of graph neural networks in real life.How to obtain a graph neural network with strong learning ability and better use of existing tags is the research objective of this paper.The paper proposes the Dual-channel Attention Mechanism Semi-supervised Graph Learning Network,which mainly includes the following three aspects:(1)Compared with most work that only explore one graph,the paper uses the original topology graph and the k nearest neighbor graph from the node features to enrich the information source of the graph neural network.While retaining the advantages of the graph neural network,it greatly enhances its ability to integrate the topology structure and node features.(2)Attention mechanism is used to guide the aggregation process.Different from the traditional graph attention network,the paper uses the similarity between node features to assign corresponding weights to different nodes in the neighborhood,so as to obtain an embedding representation containing richer semantic information.(3)The paper proposes a new semi-supervised objective function,which further uses the category information in the label data to guide the weight allocation between nodes,so as to maximize the utilization rate of label information and alleviate the dependence of semisupervised learning on label data.On the whole,the network model proposed in this paper is based on the task objective of node classification,and emphasizes the unique attribute features of nodes and the correlation among their features,showing a strong comprehensive modeling ability.In this paper,the development history of graph neural networks and classical network models at home and abroad are extensively investigated,and experiments on five public data sets show that the node classification effect of the network model proposed in this paper is superior to other comparison methods under different performance indexes,which verifies its effectiveness.In addition,visualization and ablation experiments were also used to demonstrate the advantages of the model.