基于多变量条件格兰杰因果检验的网络重构

Reconstructing Networks via Multivariate Conditional Granger Causality

摘要: 针对传统双变量格兰杰因果检验在间接因果关系识别方面效能不足，以及对滞后阶数选择敏感性过高的固有缺陷，本文创新性地构建了基于AIC准则的多变量条件格兰杰因果检验模型(LpMVGC)。该模型通过引入局部最优滞后阶数选择机制，实现了参数动态优化，有效提升了因果网络识别的准确性。该方法突破了传统检验框架的静态建模局限性，在维持计算效率的同时，增强了对复杂因果链的分析能力。本文通过引入瞬时效应视角构建了扩展型条件格兰杰因果模型，将零滞后参数纳入系统方程重构过程，有效捕捉了传统方法易忽略的瞬时因果关联。LpMVGC分别在自回归向量(VAR)模型的有向加权网络与LIF神经元有向加权网络都有良好的评价指标。

Abstract: To address the inherent limitations of traditional bivariate Granger causality testing—particularly its inefficacy in identifying indirect causal relationships and excessive sensitivity to lag-order selection—this study innovatively proposes a multivariate conditional Granger causality model (LpMVGC) based on the Akaike Information Criterion (AIC). By incorporating a locally optimal lag-order selection mechanism, the model achieves dynamic parameter optimization, significantly improving the accuracy of causal network identification. This approach overcomes the static modeling constraints of traditional frameworks while maintaining computational efficiency and enhancing the analysis capability for complex causal chains. Furthermore, we introduce an extended conditional Granger causality model from an instantaneous-effect perspective, integrating zero-lag parameters into the system equation reconstruction process to effectively capture transient causal associations often overlooked by conventional methods. The proposed LpMVGC demonstrates robust performance metrics in both vector autoregressive (VAR) model-based directed weighted networks and leaky integrate-and-fire (LIF) neuron-directed weighted networks.