Research On Node Classification Based On Graph Convolutional Networks In Heterogeneous Graphs

With the development of information technology,heterogeneous data has become the main form of information.In the real world,many heterogeneous data can be intuitively modeled as heterogeneous graphs.Node classification is an important research topic in the field of heterogeneous graph analysis.It is widely used in information recommendation,film classification,and fake review detection fields.Meta-path-based graph analysis effectively extracts information from heterogeneous graphs for node classification.At present,there are many works on node classification based on meta-path.But to my best knowledge,there is still no universal meta-path selection method.Meanwhile,nodes with a lot of neighbors and link anomalies harm node classification.To solve the above problems,this thesis propose and optimize a node classification model based on the graph convolutional network for heterogeneous graphs.The main work of this paper is as follows:(1)A node classification framework with a graph convolution network based on higher-order meta-paths is designed.It selects the meta-paths that are helpful to node classification.The framework also includes a novel heterogeneous graph convolutional network,which collects and aggregates information from different higher-order metapaths at each messaging step,and then generates node embedding for node classification.(2)This thesis propose a scheme that includes a computation method of weighted adjacency matrices based on high-order meta-paths and an optimization method for the graph convolution layer computation.To reduce the harm of nodes with a lot of neighbors,this scheme reduces the impact weight of these nodes through the calculation of the above adjacency matrices.Meanwhile,it reduces the time complexity by optimizing the graph convolution layer computation.The optimized model provides accuracy up to 1.24%?4.01%higher than competing methods,respectively.(3)A node classification method with data augmentation is constructed.To reduce the impact of link anomaly on node classification,this method predicts link probability by link predictor,adds the links that should appear,and deletes the links that should not appear according to the link probability.Through the above steps,a new denoised graph is obtained,and then it is input into this model.Experimental results show that the data enhancement helps node classification,the accuracy of this model and classical graph neural networks is improved by 0.31%?1.35%and 0.35%?1.91%respectively.Based on the graph convolutional network technology,the above three parts of the work extract meaningful information and remove noise for node classification.They help to classify nodes from three perspectives:meta-path selection,reducing the influence of nodes with a lot of neighbors,and removing the noise caused by link anomalies.Experimental results show that this model has excellent performance.