自适应频率和动态节点嵌入的图卷积网络

An adaptive frequency and dynamic node embedding based graph convolutional network

图卷积网络由于能够直接处理图结构数据的优点而受到广泛研究。当前的多数图卷积网络是基于图信号的平滑性(低频信息),且不能根据各节点适合的接受域生成对应的节点嵌入,随着网络层数的增加,易出现图卷积网络特有的过平滑问题而导致性能下降。为此,提出了基于自适应频率和动态节点嵌入的图卷积网络模型(adaptive frequency and dynamic node embedding based graph convolutional network,FDGCN)。FDGCN模型能够自适应聚合不同频率的信息,同时利用每层网络的输出,平衡每个节点来自全局和局部领域的信息,动态地调节节点嵌入。通过在4个公共数据集上进行实验,对比了6个现有模型,证明了FDGCN模型的有效性。

Graph convolutional networks have been extensively studied due to their advantages of being able to directly handle graph-structured data.Most of the current graph convolutional networks are based on the smoothness of the graph signal(low frequency information)and cannot generate corresponding node embedding according to the suitable acceptance domain of each node.However,as the number of network layers increases,the problem of over-smoothing unique to graph convolutional networks is prone to occur,resulting in performance degradation.Therefore,an adaptive frequency and dynamic node embedding based graph convolutional network(FDGCN)was proposed.FDGCN model is capable of adaptively aggregating information at different frequencies;meanwhile,it dynamically adjusts node embedding by using the output of each network layer to balance the information from the global and local domains of each node.Experiments were conducted on four public datasets comparing six existing models to demonstrate the effectiveness of the FDGCN model.