Research On Homophily And Heterophily Graph Neural Network Based On View Reconstruction

In the era of big data,how to mine the potential value from massive graph datasets with non-Euclidean distances has become a crucial issue,thus attracting the attention of a large number of researchers to study graph neural networks and graph representation learning for nonEuclidean distances.However,mainstream graph neural networks are designed based on the assumption of homophily a special form of lowpass filter that smooths the information between nodes during aggregation and achieves excellent results in homophily graphs.However,there are a large number of heterophily graphs in real-world graph data,and graph neural networks based on the homophily assumption cannot retain the high-frequency information between nodes,which makes it difficult to apply to heterophily graphs.In addition,graph neural networks based on node views cannot effectively reduce edge information indirectly from node information,resulting in poor performance in edge tasks.At the same time,graph contrast learning using graph neural networks as encoders encounters a similar problem,where the encoders for graph contrast learning are also not applicable to heterogeneous graphs.Data augmentation methods such as random edge deletion for graph contrast learning cannot effectively exploit the complex and diverse information in the graph to reduce the heterogeneity of the original graph,and cannot effectively help the encoder to learn.Therefore,the contrast learning model is not applicable to homogeneous and heterogeneous mixed graphs.The main research of the paper is to help graph neural networks and graph contrast learning models break through the assumption of homophily graphs and extend to heterophily graphs by way of view reconstruction.The main innovation points of the paper are as follows.1.To allow existing GNNs to better maintain low and high frequency signals between connected nodes and to extend to heterophily graph learning,this thesis proposes a node-pair view-based graph neural network framework(paired nodes view based graph neural network,Pair GNNs).By using node pairs as the main information aggregation unit to better preserve the information between paired nodes from being smoothed.This framework uses splicing in the intermediate process of aggregation to combine the information of its own node pairs with the aggregated information of neighboring node pairs to further ensure that the difference information between node pairs is preserved,and it is theoretically demonstrated that this design can facilitate the learning of heterogeneous graph structures.In addition,this thesis designs a neighborhood sampling method to alleviate the neighborhood explosion and reduce the computational complexity.Finally,in order to support different downstream tasks,this thesis designs an embedding transformation strategy to convert the node-to-view representation into a node-to-view representation.Three mainstream graph neural networks are chosen as variant models for this framework,and extensive experiments are conducted with seven powerful baselines under three downstream tasks and eight datasets,all of which yield excellent performance results.2.In order to solve the problem of graph contrast learning in diverse homogeneous graphs,this thesis proposes a Graph Contrastive Learning Method for Adaptive Heterophily Graphs(GCAH)that acts adaptively on diverse homogeneous graphs,and designs two data enhancement methods and a negative sample sampling method.In order to make full use of the graph structure information,this thesis uses the local homogeneity of node features to represent the structure information around the nodes,and calculates the similarity as the structure similarity weight;in order to make full use of the node feature information,the similarity of node features is calculated as the feature similarity weight.The above two weights are combined to perform adaptive structure-and feature-based data augmentation on the original graph.Then,the original graph is data augmented to produce two views,and the model is trained using contrast loss to maximize the mutual information embedded in the corresponding nodes in these two views.Finally,to avoid the effect of false negative samples,a negative sample sampling method is designed to eliminate the false negative samples similar to the positive samples in the negative samples.This framework is experimented with nine baseline models in seven datasets,and excellent results are obtained.