Bilinear Graph Neural Network With Neighbor Interactions

Graph-based data exists commonly in the real world,which indicates the relation-ships between entries.In recent years,deep learning method has show its superior per-formance in the filed of artificial intelligence.Thus,it is much potential to analyze graph-based data using deep learning method.Compared to the traditional models,such as only considering node features and graph structures,graph neural network can comprehensively explore the two kinds of information on graphs,which leads to benefit finishing the prediction tasks.Existing graph neural networks can be divided into two categories:spectral graph neural networks and spatial graph neural networks.Most traditional graph neural net-works define the convolution operation as the weighted sum of neighbor representations of the target node.However,the operation of weighted sum assumes that neighbor nodes are independent of each other.When considering such interactions,such as the co-occurrence of two neighbor nodes is a strong signal of the target node's characteris-tics,existing GNN models may fail to capture the signal.On one hand,this paper proposed a new graph convolution operator,which can explicitly model the interactions between nodes.For the proposed bilinear aggregator,the paper proved it has linear computation complexity and permutation invariant prop-erty.Based on bilinear aggregator,we design bilinear graph neural network(BGNN)by weighted aggregating multi-hop neighbor features.On the semi-supervised node classi-fication task,we test our BGNN model on three benchmark datasets.BGNN outperform GCN and GAT 1.6%and 1.5%respectively on classification accuracy.On the other hand,it's very challenging to train large GNNs.To make our BGNN scale to the large graph-based data,the paper first used METIS method to partition large graph into some small-sized graphs and then limited aggregators running on the small-sized graphs.Finally,use Pure Neighbor Aggregator method to train large scale graph neural network.In the experiments,the paper utilized large graph datasets about E-commerce and social networks.On the node classification task the paper conducted experiments and validated that the proposed methods have better performance by accu-racy and Macro-F1 than traditional GNNs while still demanding lower memory.