| 摘 要: 为了更深入地理解复杂的网络化系统,从节点的观测时间序列中挖掘非线性因果关系是一个至关重要的问题。针对此问题,提出了一种基于稀疏贝叶斯非线性条件格兰杰因果(SBNCGC)的因果网络重构的统一框架。首先,建立一个非线性条件格兰杰因果关系(NCGC)模型去捕捉目标节点与驱动节点之间的非线性关系;随后,引入稀疏贝叶斯推理方法获得目标节点的驱动节点候选集,达到了 NCGC 模型的降维作用;最后,基于重新构建的NCGC模型进行非线性格兰杰因果分析,以确定节点间因果强度。在 Kuramoto仿真数据集上,针对节点数为50的不同网络结构仿真情形,SBNCGC获得的总体平均 ROC曲线下的面积(AUROC)和 PR曲线下的面积(AUPR)值分别为87.63%和84.55%,整体上优于其他流行的格兰杰因果方法。所有仿真案例都证明了SBNCGC的优点和鲁棒性。 |
| 关键词: 格兰杰因果 网络重构 因果推理 复杂非线性系统 贝叶斯 |
|
中图分类号: TP183
文献标识码: A
|
| 基金项目: 国家自然科学基金资助项目(61903161) |
|
| Reconstruction of Complex Networks Based on Sparse Bayesian Nonlinear Conditional Granger Causality |
|
YANG Guanxue, LIN Fang
|
(School of Electrical and Information Engineering, Jiangsu University, Zhenjiang 212013, China)
gxyang@ujs.edu.cn; 2222207052@stmail.ujs.edu.cn
|
| Abstract: To gain a deeper understanding of complex networked systems, uncovering nonlinear causal relationships from observed time series of nodes is a critical issue. To address this problem, a unified framework for causal network reconstruction based on Sparse Bayesian Nonlinear Conditional Granger Causality (SBNCGC) is proposed. First, a Nonlinear Conditional Granger Causality (NCGC) model is established to capture the nonlinear relationships between target nodes and driver nodes. Subsequently, a sparse Bayesian inference method is introduced to obtain the candidate set of driver nodes for the target node, thereby achieving dimensionality reduction in the NCGC model. Finally, a nonlinear Granger causality analysis is conducted based on the reconstructed NCGC model to
determine the causal strength between nodes. In simulation datasets of Kuramoto with 50 nodes under different network structures, the SBNCGC method achieves overall average AUROC and AUPR values of 87. 63% and 84. 55% ,respectively, outperforming other popular Granger causality methods in general. All simulation cases demonstrate the advantages and robustness of SBNCGC. |
| Keywords: Granger causality network reconstruction causal inference complex nonlinear systems Bayesian |
