基于拓扑事件驱动的时序链路预测方法

Temporal Link Prediction Method Based on Topology Structure Event Driven

针对动态网络的时序链路预测方法可以基于最大似然概率模型进行扩展。一方面,该类方法大多只考虑了连边演化的情况,未考虑节点的特征影响,限制了预测性能的提升;另一方面,多数现有方法只针对动态网络的一种变化场景进行优化,没有区分不同场景的不同贡献。提出了基于拓扑事件驱动的时序链路预测方法,该方法的核心思想是把网络实体的演化过程视作一个个独立的事件过程,把节点和连边的动态变化分为新增、保持、消失等不同事件类型,先分别对连边事件和节点事件的特征进行刻画,再融合这些特征构建基于拓扑事件驱动的相似性指标,并用该指标表示连边的最大似然概率。在6个真实动态网络数据集中的实验结果表明,相比于传统经典的动态网络链路预测方法,该方法在ROC曲线下方的面积大小(Area Under Curve,AUC)和排序分数(Ranking Score,RS)2种评价标准下均取得更优的预测性能。

The temporal link prediction methods for dynamic networks can be based on the maximum likelihood probability model.On the one hand,most methods are only designed for the edge influence and do not consider the impact of the point events,which limits the improvement of prediction performance.On the other hand,these methods just optimize for the single situation of the dynamic network.This paper proposes a temporal link prediction method based on topology structure event-driven.The core idea of this method treats the evolution process of network entities as the independent events process,and divides the different event types of the appearing,maintaining,and disappearing.Then,the similarity index driven by both node and edge events is proposed,which is used to realize the maximum likelihood probability of edges.Experimental results on six real dynamic network data sets show that the proposed method achieves better prediction performance under AUC and Ranking Score evaluation than traditional classical link prediction methods.