融合编码及对抗攻击的元路径聚合图神经网络  

Metapath aggregation graph neural network for fusion coding and counterattack

作　　者：陈学刚[1] 姜征和 李佳玉 CHEN Xuegang;JIANG Zhenghe;LI Jiayu(College of Mathematics and Physics,North China Electric Power University,Beijing 102206,China;Zhizhesihai(Beijing)Technology Co.,Ltd.,Beijing 100083,China;School of Control and Computer Engineering,North China Electric Power University,Beijing 102206,China)

机构地区：[1]华北电力大学数理学院,北京102206 [2]智者四海(北京)技术有限公司,北京100083 [3]华北电力大学控制与计算机工程学院,北京102206

出　　处：《系统工程理论与实践》2024年第5期1549-1560,共12页Systems Engineering-Theory & Practice

摘　　要：异质信息网络(HIN)由于包含不同类型的节点和边,在实际问题中具有广泛的应用前景.HIN的表示学习模型旨在寻找一种有效的建模方法,将HIN中的节点表示为低维向量,并尽可能地保留网络中的异质信息.然而,现有的表示学习模型仍存在着对异质信息利用不充分的情况.为解决这一问题,本文提出了一种融合编码和对抗攻击的元路径聚合图神经网络模型(FAMAGNN),该模型由三个模块构成,分别是节点内容转换、元路径内聚合和元路径间聚合.该模型旨在解决现有HIN表示学习方法提取特征不充分的问题.同时,FAMAGNN引入了融合的元路径实例编码器,以提取HIN中丰富的结构和语义信息.此外,模型还引入了对抗训练,在模型训练过程中进行对抗攻击,以提高模型的鲁棒性.FAMAGNN在节点分类和节点聚类等下游任务中的优异表现证明了其有效性.The heterogeneous information network(HIN)has broad application prospects in practical problems due to its inclusion of different types of nodes and edges.The objective of representation learning models for HIN is to find an effective modeling method that represents the nodes in the heterogeneous information network as low-dimensional vectors while preserving the heterogeneous information in the network as much as possible.Existing representation learning models still have limitations in insufficient utilization of heterogeneous information.We propose a fusion encoding and adversarial attack meta-path aggregation graph neural network(FAMAGNN).The model consists of three module components,namely,node content transformation,intra-meta-path aggregation,and inter-meta-path aggregation,which aim to solve the problem of insufficient feature extraction in existing heterogeneous information network representation learning methods.At the same time,the model introduces a fused meta-path instance encoder to extract rich structural and semantic information in the heterogeneous information network.In addition,we introduce FGM adversarial training to perform adversarial attacks during model training to improve the robustness of the model.The outstanding performance in downstream tasks such as node classification and node clustering proves the effectiveness of this method.

关 键 词：异质信息网络 网络表示学习 元路径 对抗训练 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]