基于时序模体注意力图卷积的动态网络链路预测算法

时序网络中的动态链路预测旨在基于历史连边信息预测未来会产生的连边,是网络分析的重要组成部分,具有极大的理论研究价值和广阔的应用场景.针对现有的动态链路预测算法大多基于一阶连边关系预测未来连边,忽略了对高阶的拓扑信息和时序通联信息的挖掘和利用问题,提出一种基于时序模体注意力图卷积的动态链路预测算法.首先,提出一种时序模体邻接矩阵构建算法,利用时序模体抽取节点间的高阶拓扑和时序关系信息;然后利用隐式调节过程对网络演化过程进行建模,并使用时序模体邻接矩阵作为传输矩阵的图卷积神经网络学习节点的低维向量表示并进行迭代更新;最后以节点间表示向量作为输入,通过计算连边发生的条件密度函数值作为依据完成动态链路预测.在多个真实时序网络数据集上的实验结果表明,所提算法可有效挖掘节点间的高阶拓扑和时序信息,提高动态链路预测效果.

Dynamic link prediction algorithm based on graph convolutional networks via temporal motif-based attention

Aiming to predict edges in the future based on historical linkage status, dynamic link prediction in temporal networks is an important component of the network analysis and has great value in theoretical research and wide applications. Concerning the problem that current dynamic link prediction algorithms mostly only consider first-order relations to infer future links, while ignoring exploiting the higher-order topological and temporal relationships among nodes, this paper proposed a dynamic link prediction algorithm based on graph convolutional network via temporal motif-based attention. Firstly, it designed a temporal motif-based adjacency matrix construction algorithm, exploiting the higher-order topological and temporal relationships among nodes. Then it modeled the evolution of temporal network with latent mediation process, while iteratively updated the low-dimensional node representations with temporal motif-based adjacency matrix as the transmission matrix in graph convolutional network. Finally, it predicted the future links based on the conditional intensity function with node representations as input. Experimental results on various real-world temporal network datasets show that the proposed algorithm can effectively mine the high-order topological and temporal information among nodes, and improves the performance of the dynamic link prediction.