Abstract:

In the real world, any complex relationships can be represented as graphs, such as communication networks, biological networks, recommendation systems, etc. Link prediction is an important research topic in the field of graphs, but most of the current link prediction models are only for static graphs, and they ignore the evolution pattern of graphs in the time domain and the importance of global features in the evolution process. To this end, this paper proposes a link prediction model for dynamic graphs. First, in order to obtain high-quality global features, the model uses adversarial training to optimize the mutual information loss of global features and higher-order local features. Then a perceptual model based on a wide smooth stochastic process is used to ensure the smoothness of the global features in the time domain by constraining the mean and autocorrelation function values of the global features in the time dimension. The evolution pattern of the dynamic graph is then captured using a long and short-term memory (LSTM) network. Finally, the loss of predicted and true values is optimized using adversarial networks. The experimental results on USCB, SBM and AS datasets show that the proposed model performs well in the link prediction task of dynamic graphs, and it not only significantly improves the AUC values, but also reduces the MSE values. Also, the results of the ablation experiments show that the local features contribute to the global feature ground extraction, and the quality and smoothness of the global features play an important role in network link prediction.

Key words: dynamic graphs, link prediction, mutual information, adversarial networks, long short-term memory networks (LSTM)

摘要：

在现实世界中,任何复杂的关系都可以表示成图的形式,例如通信网络、生物网络、推荐系统等。链接预测是图领域的重要研究课题,但目前大部分的链接预测模型仅针对静态图,忽视了图在时域上的演化规律以及全局特征在演化过程中的重要性。为此,提出了一种动态图的链接预测模型。首先,为了获得高质量的全局特征,模型采用对抗训练的方式优化全局特征和高阶局部特征的互信息损失,然后利用基于宽平稳随机过程的感知模型,通过约束全局特征在时间维度上的均值和自相关函数值,以此保证全局特征在时域上的平稳性,再利用长短期记忆网络（LSTM）捕获动态图的演化规律,最后利用对抗网络优化预测值和真实值的损失。在USCB、SBM、AS数据集上的实验结果显示,本模型在动态图的链接预测任务上具有较好的表现,它不仅显著地提高了AUC值,还降低了MSE值。同时,消融实验的结果也表明,局部特征对全局特征的提取有促进作用,而且全局特征的质量和平稳性对网络链接预测有重要作用。

关键词: 动态图, 链接预测, 互信息, 对抗网络, 长短期记忆网络（LSTM）