基于多视图对比学习的动态图链接预测方法

Dynamic graph link prediction based on multi⁃view contrastive learning

链接预测旨在推断网络中缺失的边或预测未来可能出现的边.先前的链接预测研究主要集中在处理静态网络上,其目标是预测已知网络中缺失的边,然而,现实世界中许多复杂网络通常是动态变化的,使得动态网络中的链接预测任务往往比静态网络中更为复杂和困难.近年来,基于动态图表示学习的链接预测方法已经展现较好的结果,这类方法利用动态图表示学习方法学习节点的嵌入表示,以捕捉网络的结构和演化信息,从而在动态网络中实现有效的链接预测.现有方法主要采用循环神经网络或自注意力机制作为神经网络架构的组件,通过时间序列网络学习动态网络的演化信息,然而,动态网络的多样性和演化模式的可变性对基于复杂时序网络的方法提出挑战.这些方法可能很难适应不同动态网络中不断发展的演化模式,同时,在图表示学习领域,图对比学习因为其强大的自监督学习能力受到广泛关注,但是现有方法大多针对静态图,对于动态图的研究较少.为了解决上述问题,提出一种动态网络多视图对比学习的链接预测方法,不依赖额外的时序网络参数,实现动态网络的表示学习和链接预测.该方法将动态网络快照视为网络的多个视图,摆脱对比学习对数据增强的依赖.通过构建包含网络结构、节点演化以及拓扑演化三个视图的对比学习目标函数,挖掘快照内网络结构、快照间节点和网络高阶结构的演化模式学习节点表示,实现链接预测任务.最后,在多个真实数据集上进行了多类动态链接预测实验,实验结果显著优于所有基线方法,验证了所提方法的有效性.

Link prediction aims to infer missing edges in the network or predict possible future edges.Previous research on link prediction has mainly focused on dealing with static networks,to predict missing edges in known networks.However,most complex networks in the real world are dynamically changing,which often makes its link prediction more complex and difficult.In recent years,methods in link prediction based on dynamic graph representation learning have shown promising results.Such methods utilize dynamic graph representation learning methods to learn node representations to capture the structure and evolution information of the network for efficient link prediction.Existing methods mainly adopt recurrent neural network(RNN)or self⁃attention mechanism(SAM)as the components of neural network architecture,and learn the evolution information of dynamic networks through temporal networks.However,the diversity of dynamic networks and the variability of evolution patterns pose challenges to the methods based on complex temporal networks.It is difficult for these methods to adapt to the evolving evolutionary patterns in different dynamic networks.At the same time,in graph representation learning,contrastive learning has attracted extensive attention because of its powerful self⁃supervised learning ability.However,most existing methods are focused on static graphs,and few studies on dynamic graphs.To solve the above problems,this paper proposes a link prediction method based on multi⁃view contrastive learning for dynamic networks,which realizes representation learning and link prediction of dynamic networks without relying on additional temporal network parameters.Specifically,the method treats dynamic network snapshots as multiple views of the network,thereby getting rid of the dependence of contrastive learning on data augmentation.Then,we construct contrastive learning objectives including three views of network structure,node evolution,and topology evolution to mine network structure,the evolution patterns of nodes and high⁃level structure to learn node representations,ultimately realizing link prediction tasks.Finally,we conduct dynamic link prediction experiments on multiple real datasets,and the experimental results significantly outperform all the baseline methods,verifying the effectiveness of the proposed method.