基于多视图对比学习的动态图链接预测方法OA北大核心CSTPCD

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.