Causal Discovery for Cross-Sectional Data Based on Super-Structure and Divide-and-Conquer

This paper tackles a critical bottleneck in Super-Structure-based divide-and-conquer causal discovery: the high computational cost of constructing accurate Super-Structures–particularly when conditional independence (CI) tests are expensive and domain knowledge is unavailable. We propose a novel, lightweight framework that relaxes the strict requirements on Super-Structure construction while preserving the algorithmic benefits of divide-and-conquer. By integrating weakly constrained Super-Structures with efficient graph partitioning and merging strategies, our approach substantially lowers CI test overhead without sacrificing accuracy. We instantiate the framework in a concrete causal discovery algorithm and rigorously evaluate its components on synthetic data. Comprehensive experiments on Gaussian Bayesian networks, including magic-NIAB, ECOLI70, and magic-IRRI, demonstrate that our method matches or closely approximates the structural accuracy of PC and FCI while drastically reducing the number of CI tests. Further validation on the real-world China Health and Retirement Longitudinal Study (CHARLS) dataset confirms its practical applicability. Our results establish that accurate, scalable causal discovery is achievable even under minimal assumptions about the initial Super-Structure, opening new avenues for applying divide-and-conquer methods to large-scale, knowledge-scarce domains such as biomedical and social science research.

💡 Research Summary

The paper addresses a major bottleneck in super‑structure‑based divide‑and‑conquer causal discovery: the expensive construction of a highly‑recall super‑structure when conditional independence (CI) tests are costly and domain knowledge is scarce. Instead of insisting that the super‑structure contain (or closely approximate) the true skeleton, the authors propose a lightweight framework that only requires the super‑structure to be a high‑precision subgraph of the true skeleton. In other words, the scaffold may miss many true edges but any edge it does contain is highly reliable.

The framework consists of four modules. First, a super‑structure construction module builds a sparse scaffold using the Chow‑Liu algorithm on a pairwise dependence matrix. Dependence is measured with Copula entropy, a non‑parametric, model‑free metric that captures complex, non‑Gaussian relationships better than Pearson correlation. The resulting maximum‑spanning‑tree (MST) is extremely sparse, dramatically reducing the cost of the initial scaffold.

Second, a causal dividing module applies a graph‑partitioning algorithm (Girvan‑Newman edge‑betweenness) to the MST, splitting the variable set into several subgraphs. This dimensionality reduction limits the subsequent CI testing to much smaller subproblems.

Third, each subgraph is processed by a subgraph learning module that performs a two‑phase constraint‑based search: a forward phase adds edges for any untested pair that shows dependence, and a backward phase removes edges that become conditionally independent after conditioning on already learned neighbors. Because the search starts from the already‑partitioned subgraph, the number of candidate edges is far smaller than in a full‑graph PC run.

Fourth, a subgraph merging module reconciles the local skeletons into a global directed acyclic graph (DAG). Before merging, the algorithm revisits all previously untested node pairs across subgraphs, applying the same CI test to correct any d‑separation violations introduced by the weak super‑structure. The final merging follows the method of Shah et al., ensuring a coherent global graph.

The authors evaluate the approach on synthetic Gaussian Bayesian networks and on a real‑world longitudinal health survey (CHARLS). Two main experimental tracks are reported.

1. Ablation of the causal dividing (CD) module – Synthetic graphs with 20–40 nodes, 5,000 samples each, and linear Gaussian SEMs are used. When the CD module is enabled, precision, recall, F1‑score, and structural Hamming distance (SHD) all improve compared with a version that skips partitioning. CI test count rises only modestly, reflecting the extra tests needed to repair d‑separation violations caused by missing edges in the weak scaffold. The benefit grows with graph size, confirming that partitioning under a weak super‑structure yields better accuracy with limited overhead.

2. Comparison of dependence measures for scaffold construction – For 24‑node graphs under four noise distributions (Gaussian, exponential, gamma, uniform), four metrics (Copula entropy, mutual information, Pearson, Spearman) are tested. Copula entropy consistently yields the highest precision, lowest SHD, and the fewest CI tests, especially under non‑Gaussian noise, demonstrating its robustness for building reliable weak scaffolds.

On the CHARLS dataset, the proposed method achieves structural accuracy comparable to PC and FCI while reducing the number of CI tests by a large margin, confirming practical applicability in a domain‑scarce, high‑dimensional setting.

Strengths:

• Eliminates the need for an expensive, high‑recall super‑structure; the scaffold is built without any CI tests.
• Demonstrates that a high‑precision, low‑density scaffold suffices for effective divide‑and‑conquer, preserving or even improving structural accuracy.
• Introduces Copula entropy as a versatile dependence estimator that works well across heterogeneous noise models.

Limitations:

• Because the scaffold may omit many true edges, additional CI tests are sometimes required during the merging phase, which can erode some of the computational savings.
• The current scaffold is limited to a tree (MST) structure; more complex community‑based scaffolds are not explored.
• Experiments focus on linear Gaussian SEMs and a single real‑world dataset; broader validation on highly non‑linear or mixed‑type data is needed.

Future directions suggested by the authors include: (i) ensemble or adaptive refinement of weak scaffolds during learning, (ii) extending the partitioning step to handle non‑tree scaffolds (e.g., using spectral clustering or modularity‑based cuts), and (iii) testing the framework on diverse domains such as genomics, imaging, or text where variables are numerous and domain knowledge is limited.

Overall, the paper presents a compelling argument that accurate, scalable causal discovery can be achieved with minimal assumptions about the initial super‑structure, opening new avenues for applying divide‑and‑conquer strategies in large‑scale biomedical and social science research.