面向节点分类任务的节点级自适应图卷积神经网络  

Node-Level Adaptive Graph Convolutional Neural Network for Node Classification Tasks

作　　者：王鑫隆 胡睿 郭亚梁 杜航原[1] 张槟淇 王文剑[2,3] WANG Xinlong;HU Rui;GUO Yaliang;DU Hangyuan;ZHANG Binqi;WANG Wenjian(School of Computer and Information Technology,Shanxi University,Taiyuan 030006;Key Laboratory of Computational Intelligence and Chinese Information Processing of Ministry of Education,Shanxi University,Taiyuan 030006;Department of Network Security,Shanxi Police College,Tai-yuan 030401)

机构地区：[1]山西大学计算机与信息技术学院,太原030006 [2]山西大学计算智能与中文信息处理教育部重点实验室,太原030006 [3]山西警察学院网络安全保卫系,太原030401

出　　处：《模式识别与人工智能》2024年第4期287-298,共12页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.U21A20513,62076154);山西省重点研发计划项目(No.202202020101003,202302010101007);山西省基础研究计划项目(No.202303021221055)资助。

摘　　要：图神经网络通过对图中节点的递归采样与聚合以学习节点嵌入,而现有方法中节点采样与聚合的模式较固定,对局部模式的多样性捕获存在不足,从而降低模型性能.因此,文中提出节点级自适应图卷积神经网络(Node-Level Adaptive Graph Convolutional Neural Network,NA-GCN).设计基于节点重要性的采样策略,自适应地确定各节点的邻域规模.同时,提出基于自注意力机制的聚合策略,自适应地融合给定邻域内的节点信息.在多个基准图数据集上的实验表明,NA-GCN在节点分类任务上具有较优性能.Graph neural networks learn node embeddings by recursively sampling and aggregating information from nodes in a graph.However,the relatively fixed pattern of existing methods in node sampling and aggregation results in inadequate capture of local pattern diversity,thereby degrading the performance of the model.To solve this problem,a node-level adaptive graph convolutional neural network(NA-GCN)is proposed.A sampling strategy based on node importance is designed to adaptively determine the neighborhood size of each node.An aggregation strategy based on the self-attention mechanism is presented to adaptively fuse the node information within a given neighborhood.Experimental results on multiple benchmark graph datasets show the superiority of NA-GCN in node classification tasks.

关 键 词：自适应采样 自适应聚合 节点分类 图神经网络(GNNs) 谱图理论 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]