面向非稳态霍克斯过程的格兰杰因果发现算法研究

Research on Algorithms of Granger Causality Discovery for Non-Stationary Hawkes Process

发现非稳态霍克斯过程中潜在的格兰杰因果关系具有重要意义。现有的因果发现算法主要基于稳态性假设,无法适用于非稳态的情况。为此,文中提出了一种面向非稳态霍克斯过程的格兰杰因果发现算法:首先,建立非稳态霍克斯过程因果网络结构学习模型;然后,利用贪婪算法来发现某段霍克斯过程存在的模式;最后,利用基于极大似然估计的稀疏组套索(MLE-SGL)的方法发现模式对应的因果关系。在模拟数据上的实验效果验证了算法的正确性和有效性,并在交互式网络电视(IPTV)数据集上发现了一些不同模式及有趣的因果关系。

The discovery of potential Granger Causality in the non-stationary Hawkes process is of great significance. Existing causal discovery algorithms are mainly based on the assumption of stationary situation and thus cannot be applied to non-stationary situations. For this reason, this paper proposes an algorithm named Granger Causality for Non-stationary Hawkes Process (GC- NOHP): First, the work establishes the non-stationary Hawkes process causal network structure learning model;Then, the greedy algorithm is implemented to discover the patterns that are hidden in a certain Hawkes process and an MLE-SGL-based method is utilized to discover the corresponding Granger causality of the patterns. The experimental results on synthetic data verified the correctness and effectiveness of the algorithm, and the algorithm also can find different patterns and some interesting causal relationships on the Interactive Internet TV (IPTV) data set.