Mitigating Confounding Effects in Causal Discovery via Time-lagged Backdoor Pathways

Tianzhi Wang, Wenjie Yuan, Xuejuan Zhang, Shuangyan Wang, Xiaowei Zhang, Bin Hu

Keywords: causal discovery, time-lagged causality, confounding effects, Backdoor Pathways

TL;DR: A novel causal discovery algorithm for the computation of time-lagged causality under various interactions, without the assumption of no latent variables.

Abstract: Mitigating confounding effects is one of the fundamental challenges in causal discovery. This difficulty is amplified in more complex causal structures: where interactions involve colliders, mediators, and their hybrids, methods tailored for handling confounders may incur substantial errors, especially in the presence of latent factors. In this paper, we propose a novel causality discovery algorithm of Conditional Independence Test on Time-Lagged Backdoor Pathways ($\textbf{CIT-TBP}$). This approach intelligently leverages backdoor pathways induced by time-lagged causation to indirectly infer causal relationships, effectively eliminating the influence of various forms of complex interactions. Furthermore, by incorporating causal information flow, our method significantly reduces the impact of latent variables. We theoretically prove the rationality and effectiveness of the algorithm and experimentally validate it on several synthetic and real datasets. The experiment results demonstrate the superiority of our CIT-TBP against state-of-the-art methods. Compared with contemporary optimization-based methods, our causal discovery framework does not involve any black-box optimization process, and thus the causality derived are more direct and have a wide range of potential applications. The code is available at https://anonymous.4open.science/r/CIT-TBP-F0E8.

Primary Area: causal reasoning

Submission Number: 10224