Local data enhancement method of icml2022 | graph neural network

Thesis link ：https://arxiv.org/pdf/2109.03856.pdf

This article is written by Tencent AI Lab The dominant , Stanford University , Hong Kong University of science and technology , Penn State University cooperated to complete , A new data enhancement method based on conditionally generated graph neural network is proposed , It can be embedded into the modeling process of arbitrary graph neural network as a plug and play module , So as to significantly improve the performance of the model ; Applicable to drug discovery 、 E-commerce recommends 、 Social networks and other extensive application scenarios .

 

On the data and tasks of graph structure , Figure neural network (GNN) Remarkable performance has been achieved .GNN The key design idea of is to aggregate the neighborhood information of each node , Get a richer representation of the information content of this node . However , For nodes with only a few neighbors , How to effectively aggregate its neighborhood information to obtain the optimal representation , There is no final conclusion at present .

 

For this problem , This paper presents a simple and effective data enhancement method , Local data enhancement , That is, by learning the conditional probability distribution represented by the neighborhood node about the central node , Generate richer features , To enhance GNN The ability to express . Local data enhancement is a widely applicable framework , It can be plug and play embedded into any GNN In the model . This method samples additional eigenvectors about each node from the learned conditional probability , And as the expanded data for the training of the model .

 

Through a lot of experiments and analysis , We have proved that this method can improve the consistency on a variety of graph structure data and different graph Neural Networks . for instance , stay Cora,Citeseer and Pubmed On dataset , Graph convolution neural network with local data enhancement is added (GCN) And graph attention network (GAT) The average accuracy of the test can be improved respectively 3.4% And 1.6%. Besides , In large graph datasets OGB Experiments on have also proved , Compared with other features in graph 、 Methods of data enhancement at the structural level , It has better effect on the task of graph node classification .