多路动态注意力图卷积网络

A study on multiplex dynamic attention in graph convolutional network

图卷积网络在处理网络数据的各种任务上有很多应用,特别是异构图卷积网络更是得到了广泛应用.但是,现有的工作中大多只设计单一关系的异构网络,这在处理那些含有多类型节点并且节点间具有多重网络关系的复杂异构网络时,很难准确捕获到跨不同关系的节点的特征,因此,这些异构网络的效果自然是不能令人满意的.为了应对这一挑战,提出了一种处理异构网络嵌入的多路动态注意力图卷积网络(MDAGCN).MDAGCN首先通过动态聚合同一节点对之间多种关系的信息,然后通过注意力机制得到不同长度元路径注意力权重,最后使用多层卷积结合注意力权重充分获取不同长度元路径上的信息来学习节点表示.在3个真实的数据集上进行实验.结果表明:MDAGCN模型可以通过动态权重以及注意力充分学习异构网络上的节点信息,以此提升节点分类和链路预测的效果.

Graph convolutional networks had many applications in various tasks of processing network data,especially in heterogeneous graph convolutional networks.However,most existed works only designed heterogeneous networks with a single relationship,which made it difficult to accurately capture the characteristics of nodes across different relationships when dealt complex heterogeneous networks with multiple types of nodes and multiple network relationships between nodes.Therefore,the effectiveness of these heterogeneous networks was naturally unsatisfactory.To address this challenge,a multiplex dynamic attention in graph convolutional network(MDAGCN)was proposed to handle heterogeneous network embeddings,MDAGCN firstly dynamically aggregated information about multiple relationships between the same node pair,and then used attention mechanisms to obtain attention weights for different length meta paths.By combining multi-layer convolution with attention weights,information on different length meta paths was fully obtained to learn node representations.Experiments results on three real datasets demonstrated that the proposed MDAGCN model could fully learn node information on heterogeneous networks through dynamic weights and attention,thereby improving the performance of node classification and link prediction.