Neighborhood Relationship Research On Graph Neural Networks

As the typical representative of data representation,topological graph(Graph)can not only reflect the features of entities,but reflect the relationship between entities.Benefit from its strong data description ability,graph is often used to express complex data such as chemical molecular structure and social network.Graph Neural Networks(GNNs),the deep learning model based on graph,is able to aggregate node information and capture the abstract representation of entities,shining in tasks such as protein classification,knowledge graph and social network,etc.Many data mining studies have proved that it is crucial to learn the neighborhood relationship of samples to reveal the potential interaction rules between entities.However,for Euclidean data,such as audio and text,there is no topological relationship,it is necessary to define appropriate methods to mine the interaction rules between entities.How to construct the graph for Euclidean data and make it suitable for GNNs remains to be studied.Most of the existing methods directly utilize Euclidean distance to measure the feature similarity of nodes,and use the similarity to define the neighborhood relationship,which is vulnerable to the influence of outliers.This paper will propose a better executable scheme.In addition,although Non-Euclidean data,for instance social networks,has natural graph,they are often disturbed by noise,mainly manifested in the chaotic neighborhood relationship of entities,which makes the performance of GNNs plummet.The reason is that GNNs only considers the similarity of node features when performing node classification tasks,ignores the local structural features of the original graph,and does not dynamically optimize the graph.As deepening the number of network layers,the above problems will be accompanied by the over smoothing phenomenon of GNNs,making it hard to remove the noise.This paper will propose a corresponding method to solve the dynamic optimization problem of graph and alleviate the over smoothing problem of GNNs.To sum up,this paper proposes two methods to mine neighborhood relations which are fused with Graph Convolution neural Networks(GCNs)to complete the semi-supervised classification task of samples:(1)Neighborhood selection method based on node feature similarity,mainly applicable to Euclidean data.In this method,the spatial transformation matrix is introduced to transform the features,so that the neighborhood relationship mining can be carried out in the low dimensional mapping space.Benefit from the mining work and the space-optimization are executed synchronously,Euclidean distance is more suitable to measure the similarity among samples in this space.A profit by the optimization of the loss does not rely on the gradient descent algorithm,this method is also adapt for Non-Deep Learning.Experiments on 12 Euclidean datasets show that the neighborhood relationship constructed by this method is better than related methods,and the classification performance has been significantly improved.Especially,all classification accuracy on Cloud,WDBC,Iris and Thyroid exceed90%,which is unmatched by the related methods.This method provides a feasible scheme for how to construct graph for Euclidean data.(2)Neighborhood relationship construction method based on attention mechanism.To aggregate neighborhood node information,this method takes neighborhood structure similarity and node feature similarity into account.In this method,the number of neighbor nodes and association strength are adaptive by the algorithm,which means the neighborhood relationship can be constructed end-to-end.The visual analysis shows that the high-order features of similar nodes can be effectively aggregated and different types of nodes can be effectively separated.10-fold cross-validation experiments indicate the classification accuracy on the baseline data sets Cora,Citeseer and Pubmed is higher than that of GCNs and Graph Attention Networks(GAT),which proves that the neighborhood structure similarity is helpful to optimize the graph and alleviate the over smoothing problem,so that the classification performance is improved.