基于时空注意力深度模型的动态链路预测

Dynamic Link Prediction Based on Spatio-temporal Attention Deep Model

动态网络的链路预测在复杂网络的各个领域均有应用.动态网络的节点和链路随时间动态变化(出现或者消失),因此其链路预测比静态链路预测更具有挑战.本文提出一种时空注意力的深度模型(GLAT),通过提取动态网络的时空特征,实现端到端(end-to-end)的动态网络的链路预测.GLAT通过注意力长短时记忆网络(LSTM-attention)和注意力图卷积网络(GCN-attention)相结合,利用LSTM-attention学习网络节点连边状态的时序信息,利用GCN-attention学习每个时刻网络的结构特征,通过提出的两种时空注意力机制可有效关注与动态链路预测任务相关的时空特征.本文对四个真实世界的数据集展开实验验证,GLAT模型在AUC、GMAUC、误差率这几个指标上分别比对比算法提高了9. 41%、13. 76%、82. 41%.本文使用度中心性(DC)和链路介数中心性(EBC)来衡量每条链路的重要性,实验证明,GLAT模型在这两个重要性链路上的预测误差率上比对比算法分别提高了32. 2%、17. 77%.因此GLAT模型在预测准确性,错误率和动态跟踪方面优于现有方法.

Link prediction for dynamic networks is used in all areas of complex networks. The nodes and links of a dynamic network dynamically change( occur or disappear) over time,so their link prediction is more challenging than static link prediction. This paper proposes a deep model( GLAT),which realizes the link prediction of end-to-end dynamic networks by extracting the spatio-temporal features of dynamic networks. GLAT combines the attentional long-short term memory network( LSTM-attention) with the attention-based convolutional network( GCN-attention),which uses LSTM-attention to learn the temporal information of the network link state of each node,and uses GCN-attention to learn the structural feature of each snapshot. The spatio-temporal feature related to dynamic link prediction tasks can be effectively paid attention to by the proposed two spatio-temporal attention mechanisms. In this paper,four real-world datasets are experimentally verified. Compared to other baselines,the performance of GLAT model in AUC,GMUUC,and Error Rate respectively increases9. 41%,13. 76%,and 82. 41%. This paper uses degree-centricity( DC) and edge betweenness centrality( EBC) to measure the importance of each link. Experiments show that the prediction performance of the GLAT model on these two important links respectively increases 32. 2%,17. 77%. The GLAT model is superior to the existing methods in terms of prediction accuracy,error rate,and dynamic tracking.