基于稀疏注意力的自适应图池化方法

Adaptive graph pooling method based on sparse attention

针对图神经网络池化中不能充分保留图的局部特征的问题,提出一种基于稀疏注意力自适应图池化方法。首先,运用稀疏注意力自适应地选择相似度高的节点形成集群;然后,运用局部聚合卷积,通过节点聚合形成集群表示,选取集群表示最大的topk个节点完成采样;最后,在池化时保留图的局部特征以降低信息损失,从而提高图分类的性能。实验结果表明,与传统图池化方法相比,提出方法的分类正确率有所提升。

An adaptive graph pooling method based on sparse attention is proposed to solve the problem that the graph neural network pooling can not fully preserve the local features of the graph.Firstly,sparse attention is used to adaptively select nodes with high similarity to form clusters.Then,local aggregation convolution is used to form cluster representation through node aggregation.The nodes mapping to the largest top-k cluster representation are selected to complete the sampling.Finally,the local features of the graph are retained to reduce information loss and improve the performance of graph classification.Experimental results show that the proposed method has a higher classification accuracy than the prevailing graph pooling methods.