摘要
近年来,以图对比学习为代表的图自监督学习已成为图学习领域的热点研究问题,该类学习范式不依赖于节点的标签并具有良好的泛化能力。然而,大多数图自监督学习方法采用静态图结构设计学习任务,如对比图的结构学习节点级或者图级的表示等,而未考虑图随时间的动态变化信息。为此,文中提出了一种基于对比预测的自监督动态图表示学习方法(DGCP),利用对比损失引导嵌入空间捕获对预测未来图结构最有用的信息。首先,利用图神经网络对每个时间快照图编码,得到对应的节点表示矩阵;然后,使用自回归模型预测下一时间快照图中的节点表示;最后,利用对比损失和滑动窗口机制对模型进行端到端的训练。在真实图数据集上进行实验,结果表明,DGCP在链接预测任务上的表现优于基准方法。
In recent years,graph self-supervised learning represented by graph contrastive learning has become a hot research to-pic in the field of graph learning.This learning paradigm does not depend on node labels and has good generalization ability.However,most of the existing graph self-supervised learning methods use static graph structures to design learning tasks,such as learning node-level or graph-level representations based on structural contrast,without considering the dynamic information of graph over time.To address this problem,the paper proposes a self-supervised dynamic graph representation learning method based on contrastive prediction(DGCP),which utilizes a contrastive loss inducing the embedding space to capture the most useful information for predicting future graph structures.Firstly,each temporal snapshot graph is encoded using a graph neural network to obtain its corresponding node representation matrix.Then,an autoregressive model is used to predict node representations in the next temporal snapshot graph.Finally,the model is trained end-to-end by using the contrastive loss and sliding window me-chanism.Experimental results on real graph datasets show that DGCP outperforms baseline methods on the link prediction task.
作者
蒋林浦
陈可佳
JIANG Linpu;CHEN Kejia(School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;Jiangsu Key Laboratory of Big Data Security&Intelligent Processing(Nanjing University of Posts and Telecommunications),Nanjing 210023,China)
出处
《计算机科学》
CSCD
北大核心
2023年第7期207-212,共6页
Computer Science
基金
国家自然科学基金(61876091)
南京大学计算机软件新技术国家重点实验室开放课题(KFKT2022B01)
南京邮电大学校级科研基金(NY221071)。
关键词
动态图表示学习
对比学习
图神经网络
链接预测
Dynamic graph representation learning
Contrast learning
Graph neural network
Link prediction
