计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (12): 163-169.DOI: 10.3778/j.issn.1002-8331.2012-0285

• 模式识别与人工智能 • 上一篇    下一篇

动态连续时间网络表示学习

王岩,任浩,王喆   

  1. 吉林大学 计算机科学与技术学院,长春 130000
  • 出版日期:2022-06-15 发布日期:2022-06-15

Dynamic Continuous Time Network Representation Learning

WANG Yan, REN Hao, WANG Zhe   

  1. College of Computer Science and Technology, Jilin University, Changchun 130000, China
  • Online:2022-06-15 Published:2022-06-15

摘要: 随着时间的推移,网络会随着节点和连边的变化不断发展。针对传统网络表示学习算法不能正确处理动态网络的问题,提出一种基于随机游走的动态连续时间网络表示学习算法(DCTNE)。通过定义一个灵活的节点时序邻居概念,设计一个有偏的随机游走过程。根据时间信息,有效地探索节点的不同时序邻居并建模不同邻居的影响,学习网络表示。实验证明了DCTNE动态网络时序信息的有效性。在链接预测任务上,DCTNE的AUC值与其他算法相比最高获得了50%的增益;在节点分类任务上,DCTNE相较于其他算法在效果上有明显提升。结果表明,对网络中时间依赖关系进行建模有助于后续的网络分析任务。

关键词: 网络表示学习, 随机游走, 时序邻居, 连续时间, 动态网络, 网络演化

Abstract: The network would continue to evolve with changes in nodes and connections over the time. Aiming at the problem that traditional network representation learning algorithms cannot handle dynamic networks correctly, a dynamic continuous-time network representation learning algorithm based on random walks(DCTNE) is proposed. By defining a flexible node timing neighbor concept, a biased random walk process is designed. According to the time information, it can effectively explore the neighbors of different time series of nodes and model the influence of different neighbors, and learn the network representation. The experiment proves the effectiveness of DCTNE dynamic network timing information. On the link prediction task, the AUC value of DCTNE is up to 50% gain compared with other algorithms. On the node classification task, DCTNE also has significantly improved the performance. The results show that modeling the time dependence in the network is helpful for subsequent network analysis tasks.

Key words: network representation learning, random walk, temporal neighbor, continuous time, dynamic network, network evolution