基于特征选择与瞬时条件独立性检验的两阶段时间序列因果推断方法

A Two-Stage Method for Time Series Causal Inference Based on Feature Selection and Momentary Conditional Independence Testing

摘要: 从观测时间序列数据中识别并量化因果关系，是理解复杂动力系统的核心科学问题。然而，由于实际观测数据往往具有高维、非线性且样本量有限等特征，传统数据驱动方法常受特征冗余干扰，难以获得可靠的因果推断结果。为此，本文提出一种基于全局冗余最小化的自动加权特征选择框架(AGRM)与瞬时条件独立性(MCI)检验的两阶段因果推断方法。在第一阶段，首先利用信息增益(IG)计算变量间的特征得分，随后引入AGRM从全局视角剔除原始时间序列中的无关和冗余变量，从而精准提取目标变量的最相关特征集合。在第二阶段，进行MCI检验以推断变量间的因果关系，从而刻画出时间序列的因果网络结构。最后，在一个合成数据集上进行实验，并引入真阳性率、假阳性率以及运行时间，作为方法性能评价指标。结果表明，所提方法能够有效刻画变量间的因果关系，为时间序列的因果关系推断提供了一种兼顾冗余抑制与因果检验能力的有效思路。

Abstract: Identifying and quantifying causal relationships from observed time series data is a core scientific problem for understanding complex dynamical systems. However, real-world observational data are often characterized by high dimensionality, nonlinearity, and limited sample sizes, which make traditional data-driven methods vulnerable to interference from redundant features and hinder the acquisition of reliable causal inference results. To address this issue, this paper proposes a two-stage causal inference method that integrates an Automatic Global Redundancy Minimization-based feature selection framework (AGRM) with Momentary Conditional Independence (MCI) testing. In the first stage, Information Gain (IG) is first used to calculate feature scores among variables, and AGRM is then introduced to remove irrelevant and redundant variables from the original time series from a global perspective, thereby accurately extracting the most relevant feature set for the target variable. In the second stage, MCI testing is performed to infer causal relationships among variables, thus characterizing the causal network structure of the time series. Finally, experiments are conducted on a synthetic dataset, where the true positive rate, false positive rate and computational time are adopted as performance evaluation metrics. The results show that the proposed method can effectively characterize causal relationships among variables, providing an effective approach for time series causal inference that balances redundancy suppression and causal testing capability.