基于时空正则化流的动态链路预测

Temporal link prediction on dynamic graphs based on spatialtemporal normalizing flow

动态图作为图的一个重要分支,对节点间关系的动态变化过程具有良好的表达能力.利用动态图对实际关系网络进行建模,并动态预测未来时刻节点间的链路关系成为当前研究热点.然而,由于弱关系现象的存在,加权网络中的动态链路预测面临着重大挑战.针对这一问题,本文提出了一种基于正则化流的方法DynWFlow(dynamic weight flow).该方法能够从生成角度出发,自适应地评价节点间链路信息的重要性,从而精准地进行链路特征的抽取,有效地解决了动态链路预测问题.特别地,对于弱关系情况,提出利用邻居节点集权重的相似程度来评估不同链接关系的重要程度,实现对节点间隐含关系的进一步捕获.在多个领域大量真实数据的实验结果表明,所提出的基于正则化流的动态链路预测方法DynWFlow的性能明显优于其他预测算法.

Dynamic graphs,as a crucial branch of graph theory,possess comprehensive capabilities for capturing the dynamic changes in relationships between nodes.Modeling real-world relational networks using dynamic graphs and dynamically predicting the link relationships between nodes in the future have become current research hotspots.However,due to the phenomenon of weak relationships,dynamic link prediction in weighted networks faces significant challenges.Addressing this issue,this paper proposes a method based on regularization flow called DynWFlow(dynamic weight flow),which is based on normalizing flow(NF).This method,starting from a generative perspective,adaptively evaluates the importance of link information between nodes,enabling precise extraction of link features and effectively resolving the dynamic link prediction problem.Particularly,for situations involving weak relationships,it is proposed to assess the importance of different linkages by utilizing the similarity of weights in the neighbor node set,achieving further capture of implicit relationships between nodes.Experimental results with extensive real-world data from multiple domains indicate that the performance of the proposed dynamic link prediction method,DynWFlow,based on NF,is significantly superior to other prediction algorithms.