Research On Node Classification Method Based On Heterogeneous Graph Convolutional Network

Supervised classification is an important application of machine learning in processing structured and unstructured data.Traditional supervised classification methods based on attribute graphs propagate node labels on graphs that are usually composed of node attribute features in graphs,while graph neural networks perform smooth operations on the attributes of nodes,while propagating node labels on real graph topologies,Not only for isomorphic graphs but also for heterogeneous graphs.As an important data type,the demand for analysis and learning of graph data is increasingly prominent.In particular,end-to-end learning on heterogeneous graphs is currently a particularly hot research topic.Based on the above background,and drawing on the convolution operation in CNN,drawing inspiration from the research hotspot of traditional graph convolutional networks(GCN),a new graph aggregation neural network called GACN for node classification on heterogeneous graphs is proposed.The main idea of GACN is to separate the heterogeneous graph nodes to be classified without the propagation of the graph convolution layer on the basis of heterogeneous GCN,and extract the attribute characteristics and adjacency matrix of the nodes connected to them,and these nodes and The adjacency information of the nodes to be classified and input into the GCN for calculation to obtain a low-dimensional neighborhood representation of the node to be classified,and the latter is used to perform aggregation calculation with the lowdimensional connected subgraph representation to propagate the label information of the node,most end-to-end To learn the category of the node to be classified.This method of independent composition of nodes to be classified and other nodes adaptively learns the decision embedding of node categories,and for the composition of various types of nodes on heterogeneous graphs,the attention mechanism is used to learn other different types of nodes to treat classification nodes At the same time,the nodes to be classified without GCN propagation also greatly reduce the amount of parameters in GCN training and speed up the training.Experiments show that the GACN method has superiority and correctness to a certain extent.Through comparison of experimental results,it is found that the GACN model improves the AUC evaluation index of classification problems by about 5%.