一种自监督的消息传递图表示学习方法  

Self-supervised Message Passing Graph Representation Learning Method

作　　者：许珂 汤颖[1] XU Ke;TANG Ying(Zhejiang University of Technology,College of Computer Science and Technology College of Software,Hangzhou 310023,China)

机构地区：[1]浙江工业大学计算机科学与技术、软件工程学院,杭州310023

出　　处：《小型微型计算机系统》2024年第9期2196-2204,共9页Journal of Chinese Computer Systems

基　　金：浙江省自然科学基金重点项目(LZ23F020010)资助;国家自然科学基金面向项目(61972355)资助;国家自然科学基金重大项目(72192820)资助.

摘　　要：图表示学习旨在通过监督或无监督方法学习图结构数据中节点的嵌入表示.对于无标签或缺乏可靠标签的图数据集,监督方法无法正常工作,现有的无监督方法也难以学得足够准确的节点嵌入.本文提出一种自监督的消息传递图表示学习方法(SMP-GL),使用多自动编码器生成并筛选基底嵌入,基于消息传递的思想通过自监督更新节点信息来学习图节点的嵌入表示,并通过对嵌入层的覆盖更新实现多级消息传递,在不使用标签信息的情况下大幅提高了节点嵌入的准确性.论文选用四个真实世界数据集和八个先进的基线方法进行对比实验,结果表明,本文模型不仅超过了以往先进的无监督方法,而且还匹配甚至在多数任务中超过了以往先进的监督方法,能够有效应用于无标签图数据集的表示学习任务.Graph representation learning aims to learn embedding representations of nodes in graph-structured data through supervised or unsupervised methods.For graph datasets that are unlabeled or lack reliable labels,supervised methods cannot work well,and existing unsupervised methods also struggle to learn sufficiently accurate node embeddings.This paper proposes a self-supervised message passing graph representation learning method(SMP-GL)that uses multiple autoencoders to generate and filter substrate embeddings,learns the embedding representation of graph nodes by self-supervised updating of node information based on the idea of message passing,and achieves multi-level message passing by overlay updates to the embedding layers,substantially improving the accuracy of node embeddings without using label information.The paper selects four real-world datasets and eight state-of-the-art baseline methods for comparison experiments,and the results show that the model in this paper not only outperforms state-of-the-art unsupervised methods,but also matches or even outperforms state-of-the-art supervised methods in most tasks,and can be effectively used for representation learning tasks on unlabeled graph datasets.

关 键 词：图表示学习 自动编码器 自监督 消息传递 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]