An iterative conditional variable selection method for constraint-based time series causal discovery

Time series causal discovery aims to identify cause-effect relationships among variables from time series data, providing valuable insights into complex real-world scenarios. However, existing constraint-based causal discovery methods face challenges such as limited detection power, stemming from issues like dimensionality explosion and uncertainty caused by indirect paths. To address these problems, we propose a novel iterative conditional variable selection method designed for lagged, linear, and nonlinear causal discovery in time series. (1) Firstly, we block indirect information while minimizing the dimensionality of the conditioning set. Specifically, our method selects the parent set of each target variable as the conditioning set, which includes only those variables involved in the indirect path. (2) Then, we refine the conditioning set by selecting a subset of the parent set for each target variable to focus on indirect causal relationships. (3) Finally, the iterative application of steps (1) and (2) progressively corrects the indirect paths, leading to a significant improvement in detection power. Experimental results on synthetic and public datasets, as well as for varying time lags, node counts, and a chemical fault diagnosis case, demonstrate that our method outperforms state-of-the-art (SOTA) approaches.