Journal of Nanjing University(Natural Sciences), Volume. 61, Issue 4, 624(2025)

A significance⁃weighted divide⁃and⁃conquer approach for causal discovery

Bai Tianxu1, Zhai Yanhui1,2, and Li Deyu1,2、*
Author Affiliations
  • 1School of Computer and Information Technology, Shanxi University, Taiyuan, 030006, China
  • 2Key Laboratory of Computational Intelligence and Chinese Information Processing of Ministry of Education, School of Computer and Information Technology, Shanxi University, Taiyuan, 030006, China
  • show less

    To address the challenges of causal discovery in high-dimensional data, this paper proposes a significance-weighted divide-and-conquer causal discovery approach (SWCD). Traditional causal discovery methods suffer from high computational complexity, ambiguous differentiation of Markov equivalence classes, and crude conflict resolution mechanisms in high-dimensional scenarios. To overcome these limitations, this work integrates divide-and-conquer strategies with significance weighting to refine the causal discovery process. Specifically, the method achieves a synergistic optimization of efficiency and accuracy through a three-tiered design. In the partitioning phase, path significance values (PSV) and path importance scores(PIS) are defined to dynamically quantify the statistical reliability of causal paths. By combining topological features and adaptive partitioning strategies, the framework prioritizes retaining high-confidence causal chains to protect critical structures while dynamically optimizing decomposition paths. In the solving phase, the PC algorithm is enhanced using residual-based conditional independence testing (ReCIT), which distinguishes Markov equivalence classes by analyzing the independence of regression residuals. In the merging phase, a confidence score-driven conflict resolution mechanism is designed to resolve edge conflicts during subgraph merging, where edge reliability is quantified through confidence scores. Experimental results demonstrate that the proposed method significantly outperforms existing baseline approaches such as CPBG on high-dimensional datasets, achieving superior performance in efficiency, robustness, and interpretability. Future research can focus on optimizing significance quantification metrics, refining dynamic partitioning strategies, and exploring adaptability to nonlinear causal relationships.

    Tools

    Get Citation

    Copy Citation Text

    Bai Tianxu, Zhai Yanhui, Li Deyu. A significance⁃weighted divide⁃and⁃conquer approach for causal discovery[J]. Journal of Nanjing University(Natural Sciences), 2025, 61(4): 624

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Received: May. 20, 2025

    Accepted: Aug. 22, 2025

    Published Online: Aug. 22, 2025

    The Author Email: Li Deyu (lidysxu@163.com)

    DOI:10.13232/j.cnki.jnju.2025.04.008

    Topics