注意力融合机制和拓扑关系挖掘的异构图神经网络  

Heterogeneous graph neural network based on attention fused mechanisms and topology relation mining

作　　者：陈金杰 王一蕾[1] 傅仰耿[1] CHEN Jinjie;WANG Yilei;FU Yanggeng(College of Computer and Data Science,Fuzhou University,Fuzhou,Fujian 350108,China)

机构地区：[1]福州大学计算机与大数据学院,福建福州350108

出　　处：《福州大学学报(自然科学版)》2025年第1期1-9,共9页Journal of Fuzhou University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(12271098);福建省高校产学合作科技计划资助项目(2023H6008)。

摘　　要：针对异构图神经网络模型依赖元路径和复杂聚合操作导致元路径受限与高成本的不足,提出一种基于注意力融合机制和拓扑关系挖掘的异构图神经网络模型(FTHGNN).该模型首先使用一种轻量级的注意力融合机制,融合全局关系信息和局部节点信息,以较低的时空开销实现更有效的消息聚合;接着使用一种无需先验知识的拓扑关系挖掘方法替代元路径方法,挖掘图上的高阶邻居关系,并引入对比学习捕获图上的高阶语义信息;最后,在4个广泛使用的现实世界异构图数据集上进行的充分实验,验证了FTHGNN简单而高效,在分类预测准确率上超越了绝大多数现有模型.Aiming at the shortcomings of heterogeneous graph neural network models,such as reliance on meta-paths and complex aggregation operations,which lead to limited meta-paths and high costs,a fused attention mechanism heterogeneous-based graph neural network model(FTHGNN)and topological relation mining is proposed.The attention fused mechanisms integrate global relationship and local node information to achieve more effective message aggregation with lower computational cost.Subsequently,a topology relation mining method that requires no prior knowledge is employed to replace the meta-path based method,enabling the mining of high-order neighbor relationships.Contrastive learning is introduced to capture high-order semantic information on the graph.Finally,extensive experiments on four widely used real-world heterogeneous graphs demonstrate that FTHGNN is simple yet efficient and outperforms most existing models in classification prediction accuracy.

关 键 词：图神经网络 异构图 注意力机制 对比学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]