基于图神经网络的多维信息拓扑结构挖掘方法  

Mining Method of Multi-dimensional Information Topology Based on Graph Neural Network

作　　者：陈杨 司毅 王吉刚 杨威 CHEN Yang;SI Yi;WANG Jigang;YANG Wei(The 8th Research Academy of CSSC,Yangzhou 225101,China)

机构地区：[1]中国船舶集团有限公司第八研究院,江苏扬州225101

出　　处：《舰船电子对抗》2025年第1期77-81,共5页Shipboard Electronic Countermeasure

摘　　要：大量的非欧几里得数据充斥在我们的日常生活中,这些繁杂的实体集合以及实体之间的联系,难以使用现有的数据模型刻画,但是能够用图结构来高效、精准地建模。图结构能够很好地建模现实生活中复杂的关联关系,图表示方法可以有效挖掘复杂关系中的模式匹配信息。在图神经网络的基础上,结合拓扑结构中的度特征和Graphlets子图特征,高效、精准地建模实体与实体之间的多维度联系。引入了预训练机制,通过在预训练模型上的微调或者知识蒸馏,从而完成实体识别分类的任务。理论与实验表明,在多维数据集中,基于图神经网络并结合图拓扑结构信息的挖掘是可行且有效的,能够提升分类任务的性能。A large number of non-Euclidean data fill our daily life.These complex entity sets and the relationships among entities are difficult to depict by using existing data models,but they can be modeled efficiently and accurately by using graph structures.Graph structure can model the complex relationship in real life well,and graph representation method can effectively mine the pattern matching information in the complex relationship.On the basis of graph neural network,combined with the degree feature in topology and the subgraph feature of Graphlets,the multi-dimensional relationship between entities can be modeled efficiently and accurately.The pre-training mechanism is introduced to complete the task of entity recognition and classification by fine-tuning or knowledge distillation on the pre-training model.The theory and experiment show that the mining based on graph neural network combined with graph topological structure information is feasible and effective in multidimensional data sets,which can improve the performance of classification tasks.

关 键 词：图神经网络 拓扑结构 预训练模型 知识蒸馏模型 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]