acceptodds
Under review as a conference paper at ICLR 2027

Decoupled Causal Discovery

Abstract

Causal discovery from observational data is a fundamental yet challenging task in scientific research. While existing approaches are primarily based on conditional independence tests, structure scores, or restrictive functional assumptions, we propose Decoupled Causal Discovery (DCD), a novel decoupling-based perspective that does not rely on these methodologies. DCD directly identifies the Markov boundary (MB) by decoupling non-target variables via weighting functions, such that only variables within the MB preserve dependence with the target under the decoupled distribution. Building on this, DCD iteratively constructs the Completed Partially Directed Acyclic Graph (CPDAG) by exploiting structural asymmetries within the MBs. We establish the theoretical identifiability, soundness, and completeness of DCD. Empirical evaluations demonstrate that DCD achieves strong performance, particularly excelling in challenging noise regimes.

Primary area:
causal reasoning
Keywords:
causal discovery, Markov boundary, decoupled distribution

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ywLnOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores a new perspective on causal discovery. Conventional causal discovery methods are primarily based on conditional independence tests (CITs), score optimization, or assumptions about specific functional causal models. In contrast, this paper proposes a new causal discovery method based on the idea of artificially decoupling dependencies among variables through sample reweighting. The proposed method first selects a target variable and assigns sample weights to the remaining variables so that they become mutually independent. Variables that remain dependent on the target variable after reweighting are then identified as belonging to the target variable’s Markov Boundary (MB). The method subsequently exploits the structural asymmetry in how the MB changes for parent/child variables versus spouse variables to determine causal relationships among the nodes.

The paper theoretically establishes the validity of MB identification under the assumptions of Strict Positivity and Weighting Faithfulness. It further shows that Weighting Faithfulness holds almost everywhere over an appropriate weight space. To learn the sample weights in practice, the method minimizes the sum of weighted HSIC values over all pairs of non-target variables. Experiments on synthetic data demonstrate that the proposed method can recover causal structures more robustly than several representative existing causal discovery methods, particularly in settings with high levels of noise.

Critical strengths

The primary strength of this paper is that it proposes a novel approach to causal discovery based on artificially decoupling dependencies among variables through reweighting, which is fundamentally different from major existing causal discovery paradigms based on conditional independence tests (CITs), score optimization, or assumptions about specific functional causal models. The paper further provides a theoretical foundation for its key MB identification mechanism, showing that it is valid under the assumptions of Strict Positivity and Weighting Faithfulness. This theoretical analysis substantially strengthens the main methodological contribution.

Another important strength is the empirical demonstration that the proposed method achieves more robust causal structure estimation than representative existing causal discovery methods, particularly under high-noise conditions. This provides useful empirical evidence of a distinctive practical advantage of the proposed approach.

Critical weaknesses

The proposed causal discovery perspective is highly interesting, and the paper is commendable for establishing identifiability under two explicit assumptions. However, a major limitation is that the practical optimization of the weights used to make the non-target variables mutually independent relies on the heuristic minimization of the sum of pairwise weighted HSIC values. The paper does not sufficiently analyze how well this optimization actually achieves the required mutual independence among the non-target variables, nor how deviations from ideal decoupling affect Markov boundary identification and, ultimately, causal discovery accuracy. A deeper analysis of this connection would substantially improve the clarity and technical completeness of the paper.

In addition, I understand that the main contribution of this work is to introduce a fundamentally different perspective from major existing causal discovery approaches. However, the motivation remains somewhat unclear in terms of what specific limitations of existing causal discovery methods the proposed approach is intended to overcome, and which components of the proposed method are responsible for these advantages. The experiments empirically show that the method is particularly robust in high-noise settings, but the paper does not clearly explain why the proposed method exhibits this robustness or which aspect of the method contributes to it. A deeper discussion of these points would help clarify the practical regimes in which this interesting approach should be preferred over existing methods.

Questions for authors
  • Can the authors discuss how much the use of minimizing the sum of weighted HSIC values affects performance compared with the ideal case in which optimal decoupling weights can be obtained?

  • Can the authors discuss which components of the proposed method address which limitations of existing causal discovery approaches, particularly how and why these components contribute to robustness in high-noise settings?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ywLn

We sincerely appreciate your recognition of our novel causal discovery perspective and theoretical foundation. Below, we provide detailed responses to your comments. If anything remains unclear or if you have further questions, please do not hesitate to let us know.

W1 / Q1. Discussion on Optimal Decoupling

Thank you for this insightful question. At the population level, for a non-degenerate continuous distribution, there can be infinitely many weight functions that satisfy the decoupling condition. However, at the finite-sample estimation level, we do not have access to the ground truth of the decoupling weights and can only perform unsupervised learning based on statistical objectives. Even if certain metrics of distributional decoupling are optimized to their optimum, these metrics are still sample-based estimators, which means their corresponding true population values fail to guarantee simultaneous optimality (even when the estimators are unbiased). Therefore, it is inherently difficult to directly compare against the optimal decoupling weights given a finite dataset.

To further investigate this, we conducted an additional comparative experiment by optimizing the joint independence objective (dHSIC) of the weighted data (under the same setting as Section 4.1 Func, d=10d=10, with all optimization and testing hyperparameters fixed; added in the revised manuscript, Appendix C.1):

MethodSHD$_c$ ↓Precision ↑Recall ↑F1 ↑
DCD5.5 ± 1.70.87 ± 0.160.45 ± 0.160.58 ± 0.16
DCD (+dHSIC loss)6.6 ± 2.10.63 ± 0.230.50 ± 0.200.55 ± 0.20

As shown above, the performance gap is not substantial, indicating that our decoupling strategy remains highly effective.

In summary, while different finite-sample estimation methods can serve our decoupling objective, it remains challenging to measure how close an empirical estimate is to the true population-level optimal decoupling. How to better learn decoupling weights in practice is a crucial and promising future research direction, not only for our method, but also for related fields in causal inference.

W2 / Q2 Discussion on High-Noise Robustness

Thank you for raising this important question. We believe the robustness of our method to high noise stems from the following estimation dynamics:

First, suppose we were given an ideal set of sample weights with a high effective sample size (ESS) that perfectly decouples the true population distribution. In that case, MB identification would become a very straightforward problem. Consequently, most of the estimation error in our current method likely arises from the imperfections in the learned decoupling weights. Specifically (as discussed in the Sample selection paragraph of Appendix C / revised PDF Appendix D), learning decoupling weights is essentially performing sample selection. This reduces the effective sample size and can lead to potential information loss over the support, ultimately introducing estimation error.

However, noise can actually be beneficial in mitigating this specific issue (as discussed in the Noise paragraph of Appendix C / revised PDF Appendix D). An intuitive example is a nearly deterministic system: the probability mass of the decoupled distribution collapses almost onto a single point, causing high-weight samples to cluster tightly together and fail to provide sufficient variation to detect dependencies. In contrast, higher noise spreads the distribution more evenly across the space. As a result, the learned decoupled distribution covers a broader support, making genuine dependencies much easier to detect.

Therefore, while higher noise generally makes standard statistical estimation harder for most methods, it actually facilitates the decoupling weight learning phase in our framework. The interplay between these two competing effects explains why our method demonstrates effectiveness under high-noise regimes.

Thank you for your detailed response and for conducting the additional experiments using dHSIC. This result has partially addressed my concerns. I also appreciate the explanation of the challenges involved in directly learning from a finite sample. I recommend that the authors incorporate these additional experiments into the paper and discuss the difficulty of evaluating the quality of the learned decoupling weights. With these additions, I believe my primary concerns will be sufficiently resolved. As I was already leaning toward acceptance, I will maintain my current rating.

Thank you for your appreciation. We have incorporated these additional experiments into the revised manuscript (Appendix C.1), and added the discussion of the difficulty (Appendix D).

Thank you again for your time and valuable feedback, which has significantly enhanced our manuscript.

Reviewer_2HBHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a decoupling-based method for causal discovery to obtain a CPDAG. The method first identifies Markov boundaries by reweighting samples so that all non-target variables are mutually independent, under which a variable is in the MB if and only if it remains dependent on the target. Spouses are then distinguished from parents and children by removing subsets of variables from the MBs and re-evaluating them, which also identifies v-structures. Simulations on two data-generating mechanisms and two graph types show that the proposed method outperforms the baselines in most settings, especially under high noise.

Critical strengths
  1. The paper is well written and easy to follow.
  2. The experiments show that the proposed algorithm outperforms the baselines in most settings, with impressive performance differences under both the functional and MLP mechanisms.
  3. The proposed algorithm is sound and complete if the model can correctly identify MBs.
  4. The computational complexity is provided in the appendix.
Critical weaknesses
  1. Regarding the theoretical section, the novelty is limited. Theorem 1 seems to be the main theoretical contribution, while the remaining results are relatively standard. Theorem 2 does not require decoupling and assumes that the correct MB is given. The proposed reweighting method requires all non-target variables to be mutually independent, which is not easy to achieve. Therefore, its advantages over standard CI tests are unclear.
  2. As for the experimental section, the evaluation is not comprehensive. Table 1 uses noise with a standard deviation of 2, a high-noise regime in which DCD outperforms the baselines. It would be more comprehensive to include other regimes, such as linear models. Furthermore, the Barabási–Albert model used in this section produces trees with small MBs, which is an ideal setting for local MB-based algorithms.
  3. Most baselines are not designed for the nonlinear settings considered, with PC using KCIT-based tests and GraNDAG being the only baselines capable of handling nonlinearity. Standard nonlinear causal discovery methods and MB-based structure learning methods, such as CAM, NoGAM, and Grow-Shrink with KCI, are missing from the comparison
  4. The advantages and motivation of the proposed algorithm could be better emphasized, particularly in comparison with CI-based algorithms.
Questions for authors
  1. It is suggested to run Algorithm 1 with a CI-based MB identification method to isolate the contribution of decoupling, as Theorem 2 holds for any MB oracle.
  2. How are the baselines’ hyperparameters tuned?
  3. A weighted correlation test is used to determine whether a variable should be included in the estimated MB. However, the weighted Pearson correlation test only detects linear dependencies. The experimental results show that the algorithm performs well in nonlinear settings, making the reason behind this performance worth exploring.
  4. How is the subset to be excluded from the obtained MB determined when identifying spouses?
  5. It is suggested to report a comparison of running times.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 2HBH (Part 1/2)

We thank the reviewer for insightful comments regarding the novelty of our method and the experimental setup. Below, we provide detailed responses to your comments. If anything remains unclear or if you have further questions, please do not hesitate to let us know.

W1 / W4 / Q1. Novelty, Theory, Advantages over CI-based methods and Motivation

Regarding the theoretical aspect, the proof techniques employed in this paper may be relatively standard. However, we would like to clarify that the core contribution of our work lies in introducing a novel perspective, with theory primarily serving to establish the framework and provide support. Theorem 2 assumes an oracle MB, which we believe is just as reasonable as the common assumption of an oracle CIT in many methods. In DCD, MB identification is formulated as a direct statistical estimation. Therefore, at the level of algorithmically invoking statistical tools, MB inference and conditional independence inference hold an analogous status.

Following your suggestion, we also added experimental comparisons where the MB estimation in DCD is replaced by CIT-based MB algorithms: GS [1], IA [2], and Pellet & Elisseeff (TC [3]), each equipped with FisherZ and KCIT (under the same setting as Section 4.1 Func, d=20d=20; added in the revised manuscript, Appendix C.3):

MethodSHD$_c$ ↓Precision ↑Recall ↑F1 ↑
DCD14.5 ± 3.40.66 ± 0.130.45 ± 0.130.53 ± 0.13
DCD-GS(FisherZ)28.5 ± 5.40.13 ± 0.180.15 ± 0.200.14 ± 0.19
DCD-GS(KCIT)20.8 ± 6.30.26 ± 0.290.27 ± 0.310.26 ± 0.30
DCD-IA(FisherZ)27.4 ± 5.80.17 ± 0.200.22 ± 0.250.19 ± 0.22
DCD-IA(KCIT)21.4 ± 5.60.23 ± 0.260.25 ± 0.280.24 ± 0.27
DCD-TC(FisherZ)25.4 ± 6.80.21 ± 0.240.23 ± 0.250.22 ± 0.25
DCD-TC(KCIT)19.7 ± 4.80.22 ± 0.240.18 ± 0.190.20 ± 0.21

Our method demonstrates clear performance advantages, highlighting the superiority of our decoupled formulation as a direct MB estimator. CIT-based methods for MB estimation exhibit substantially high variance. This occurs because multiple CIT calls are invoked during each MB identification step, causing both Type I and Type II errors of the CITs to accumulate rapidly.

Regarding the motivation of our method, thank you for your suggestion. We currently aim to highlight a novel methodological perspective. The comparison with CIT-based MB methods you suggested has been added to Appendix C.3 of the revised PDF. We would also greatly welcome any further specific suggestions regarding the paper's organization, which is valuable for improving our manuscript.

W2. Experiment Setting

We would like to clarify that, first, the functional mechanisms evaluated in Table 1 already include linear functions. Second, regarding the noise and graph structure settings, Sections 4.2 and 4.3 are dedicated to these analyses: Figure 1 presents line plots illustrating performance across various noise types ranging from near-zero to extreme levels, while Figure 2 shows results across diverse graph topologies and densities.

W3 / Q2. Baselines

Our selection of baselines was primarily guided by the goal of covering representative methods across various methodological paradigms. For the comparison with GS (KCIT) that you suggested, please refer to our first response (W1 / W4 / Q1).

Regarding implementation and hyperparameter tuning: for GES, ICALiNGAM, DirectLiNGAM, Notears, GraNDAG, and GOLEM, we use the implementations provided by gCastle [4] with their default hyperparameter settings. PC with KCIT/FisherZ is implemented using causal-learn [5], while E-KCIT/E-FisherZ is evaluated using its official implementation with default configurations [6]. Thank you for raising this point. We have also added these implementation details to the beginning of Appendix C in the revised PDF.

AuthorsOpenReview ↗

Response to Reviewer 2HBH (Part 2/2)

Q3. Implementation

Thank you for this insightful question. We believe there are two primary reasons why the weighted correlation test achieves effective empirical performance and why we use it:

  • ESS & Small-Sample Behavior: Learning decoupling weights acts similarly to sample selection, the better the decoupling, the smaller the effective sample size (ESS). In regimes with a limited ESS, complex nonparametric tests do not necessarily outperform simpler parametric ones:
    • Null Calibration: Complex nonparametric tests can suffer from poor asymptotic uniformity under the null hypothesis (H0​) in small samples, leading to inflated FDR.
    • Power: Simpler tests may achieve higher statistical power when the effective sample size is small.
  • Symmetry Breaking via Reweighting: Linear correlation tests typically lose power only under strictly symmetric nonlinear dependencies (whereas asymmetric nonlinear dependencies still yield non-zero linear correlation). However, the reweighted distribution induced by the learned decoupling weights is highly unlikely to preserve the exact symmetry of the original distribution or accidentally exhibit symmetry after reweighting.

We also added an empirical comparison with the weighted kernel (HSIC) test under the same setting as Section 4.1 (Func, d=10d=10, with all optimization and testing hyperparameters fixed; added in the revised manuscript, Appendix C.1):

MethodSHD$_c$ ↓Precision ↑Recall ↑F1 ↑
DCD5.5 ± 1.70.87 ± 0.160.45 ± 0.160.58 ± 0.16
DCD (+HSIC test)6.0 ± 2.30.67 ± 0.270.49 ± 0.230.56 ± 0.24

Overall, the performance remains quite comparable across these variants and aligns closely with our analysis: while the HSIC test can capture a broader class of dependencies (slightly improving Recall), its weaker Type I error control in small effective sample regimes leads to a drop in Precision.

Q4. How is the subset to be excluded from the obtained MB determined when identifying spouses?

When identifying spouses, we need to exhaustively iterate through all subsets of the obtained MB (we provide a detailed time complexity discussion in Appendix B). In our current implementation, this is performed via a straightforward enumeration in increasing order of subset size.

Q5. Running Times

Thank you for this suggestion. We have added a runtime comparison experiment for DCD in Appendix C.2 of the revised PDF. Below is a summary of the results under the same setting as Section 4.1 Func, evaluated on a single CPU core (methods with negligible computational costs are omitted):

Time (s)d=5d=5d=10d=10d=20d=20
DCD29.3 ± 5.5139.1 ± 21.4821.0 ± 189.8
PC (KCIT)7.5 ± 7.558.0 ± 52.1448.3 ± 830.1
GraNDAG13.5 ± 0.519.0 ± 0.232.2 ± 0.3
GOLEM29.2 ± 0.632.7 ± 0.649.6 ± 3.7

Although the computational cost is relatively higher in high-dimensional settings (primarily stemming from the optimization overhead), it remains manageable overall. Furthermore, DCD-C has potential for further parallelization (as discussed in Appendix B).


[1] Dimitris Margaritis, Sebastian Thrun. Bayesian Network Induction via Local Neighborhoods. NIPS, 1999.

[2] Ioannis Tsamardinos, Constantin F. Aliferis, and Alexander Statnikov. Algorithms for Large Scale Markov Blanket Discovery. FLAIRS, 2003.

[3] Jean-Philippe Pellet, André Elisseeff. Using Markov Blankets for Causal Structure Learning. JMLR, 2008.

[4] Keli Zhang, Shengyu Zhu, Marcus Kalander, Ignavier Ng, Junjian Ye, Zhitang Chen, and Lujia Pan. gcastle: A python toolbox for causal discovery. ArXiv, 2021.

[5] Yujia Zheng, Biwei Huang, Wei Chen, Joseph Ramsey, Mingming Gong, Ruichu Cai, Shohei Shimizu, Peter Spirtes, and Kun Zhang. Causal-learn: Causal discovery in python. JMLR, 2024.

[6] Zhengkang Guan and Kun Kuang. Efficient ensemble conditional independence test framework for causal discovery. ICLR, 2026.

AuthorsOpenReview ↗

Looking Forward to Your Feedback

Dear Reviewer 2HBH,

As the discussion period is coming to a close, we are kindly wondering if any questions or concerns remain unresolved. We would be very happy to provide any further clarification.

Thank you again for your time and consideration.

Thank you for the responses and the additional experiments!

However, my concern about the experimental setting remains. The functional mechanisms mix linear and nonlinear functions, so they do not test a purely linear setting. The main experiments, including the new CIT-based comparison, use a high-noise setting where the noise variance is four times the signal variance, and Figure 1 shows that PC with KCIT performs better under low noise. Section 4.3 also shows that PC and GES perform better on ER graphs. These results suggest that the main experiments (high noise and BA trees) favor the proposed method. It would be more convincing to show the comparison with CIT-based MB methods under lower noise levels and on linear models, and to include nonlinear baselines such as CAM and NoGAM. I therefore keep my score.

Reviewer_oTF3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper identifies Markov boundaries by reweighting samples so that the non-target variables become mutually independent. Under this reweighted distribution, a variable belongs to the target's MB if and only if it is still dependent on the target. In practice, the weights are learned by minimizing pairwise weighted HSIC, and MB membership is tested with a weighted Pearson correlation and a t-test. An MB-based algorithm, DCD-C, then exploits the asymmetry between neighbours and spouses to recover the CPDAG; it is sound and complete given an oracle MB. Experiments on synthetic Func and MLP mechanisms with d ∈ {5, 10, 20} compare against PC, GES, LiNGAM variants, NOTEARS, GraN-DAG and GOLEM.

Critical strengths
  1. The connection between stable-learning reweighting and MB discovery is interesting, and turning conditional independence into unconditional dependence under a reweighted distribution is a neat idea.
  2. Thm 1 is clean and correct under its assumptions.
  3. The paper is clearly written, and DCD has the highest recall and F1 among the compared methods in most settings.
Critical weaknesses
  1. The implementation does not satisfy the theory, in two places: Thm 1 needs mutual independence, but only pairwise HSIC is minimized. The authors acknowledge this as an approximation, but the gap matters. The sufficiency proof needs Xi to be independent of X_MB as a block, and pairwise independence does not imply that. Addtionally, Thm 1 is about general dependence, but the test is a Pearson correlation. It can miss symmetric nonlinear effects such as x² and log|x|, which appear in the paper's own mechanisms, especially since root noises are symmetric.
  2. I dont think it is "Orthogonal to CITs". Since it is still decided by an independence test, only an unconditional one. DCD-C is close to existing MB-based algorithms. These are cited but none is used as a baseline. Please state precisely what is new.
Questions for authors
  1. What happens with a joint independence objective (e.g. dHSIC) and a weighted kernel test instead of Pearson?
  2. What does DCD-C add over GS / Pellet & Elisseeff when it is given the same MBs?
  3. How does the ESS behave as d grows, what threshold is used, and what are the runtimes?
  4. How does performance scale beyond d = 20, and at lower noise levels?
4 comments
AuthorsOpenReview ↗

Response to Reviewer oTF3 (Part 1/3)

We thank the reviewer for insightful questions regarding the underlying rationale and novelty of our method. Below, we provide detailed responses to your comments. If anything remains unclear or if you have further questions, please do not hesitate to let us know.

W2 / Q2. Novelty, "Orthogonal to CITs", and Comparison with CIT-based MB Methods

We thank the reviewer for pointing out this important nuance. Strictly speaking, marginal independence is a special case of conditional independence with an empty conditioning set. However, our use of the term "orthogonal to CITs" is intended to emphasize the contrast with constraint-based methods that inherently rely on CITs with non-empty conditioning sets. When conditioning sets are present, we argue that they exhibit fundamentally different properties and difficulties compared to marginal independence:

  • Testability: It has been proven that no single CIT method is effective across all conditional structures with non-empty conditioning sets [1], whereas marginal independence testing is far more tractable and well-studied.
  • Dimensionality: Testing with conditioning sets transforms the problem from a fixed 2-dimensional marginal problem into a multi-dimensional one, leading to the curse of dimensionality. (Note: this does not mean DCD is entirely free from high-dimensional challenges; see our detailed discussion in Appendix C of the paper.)

More importantly, DCD offers a fundamentally different paradigm: it strictly does not perform CITs (with conditioning sets). From our decoupled perspective, MB estimation is cast as a direct statistical estimation problem rather than an algorithm relying on CITs as statistical tools (e.g., GS [2], Pellet & Elisseeff [4]). (Specifically, although Algorithm 3 in Pellet & Elisseeff [4] is formulated as an MB estimator, its underlying Algorithm 2 for identifying spouses and the final experiments in their paper are entirely based on CITs.)

Following your suggestion, we also added experimental comparisons where the MB estimation in DCD is replaced by CIT-based MB algorithms: GS [2], IA [3], and Pellet & Elisseeff (TC [4]), each equipped with FisherZ and KCIT (under the same setting as Section 4.1 Func, d=20d=20; added in the revised manuscript, Appendix C.3):

MethodSHD$_c$ ↓Precision ↑Recall ↑F1 ↑
DCD14.5 ± 3.40.66 ± 0.130.45 ± 0.130.53 ± 0.13
DCD-GS(FisherZ)28.5 ± 5.40.13 ± 0.180.15 ± 0.200.14 ± 0.19
DCD-GS(KCIT)20.8 ± 6.30.26 ± 0.290.27 ± 0.310.26 ± 0.30
DCD-IA(FisherZ)27.4 ± 5.80.17 ± 0.200.22 ± 0.250.19 ± 0.22
DCD-IA(KCIT)21.4 ± 5.60.23 ± 0.260.25 ± 0.280.24 ± 0.27
DCD-TC(FisherZ)25.4 ± 6.80.21 ± 0.240.23 ± 0.250.22 ± 0.25
DCD-TC(KCIT)19.7 ± 4.80.22 ± 0.240.18 ± 0.190.20 ± 0.21

Our method demonstrates clear performance advantages, highlighting the superiority of our decoupled formulation as a direct MB estimator. CIT-based methods for MB estimation exhibit substantially high variance. This occurs because multiple CIT calls are invoked during each MB identification step, causing both Type I and Type II errors of the CITs to accumulate rapidly.

AuthorsOpenReview ↗

Response to Reviewer oTF3 (Part 2/3)

W1 / Q1. Implementation

Following your suggestion, we added empirical comparisons utilizing the joint independence objective (dHSIC) and the weighted kernel (HSIC) test under the same setting as Section 4.1 (Func, d=10d=10, with all optimization and testing hyperparameters fixed; added in the revised manuscript, Appendix C.1):

MethodSHD$_c$ ↓Precision ↑Recall ↑F1 ↑
DCD5.5 ± 1.70.87 ± 0.160.45 ± 0.160.58 ± 0.16
DCD (+dHSIC loss)6.6 ± 2.10.63 ± 0.230.50 ± 0.200.55 ± 0.20
DCD (+HSIC test)6.0 ± 2.30.67 ± 0.270.49 ± 0.230.56 ± 0.24
DCD (+dHSIC loss, +HSIC test)6.1 ± 2.90.57 ± 0.310.54 ± 0.320.54 ± 0.30

Overall, the performance remains quite comparable across these variants. Specifically, guided by the principle of keeping the estimation procedure as simple as possible, our implementation choices are motivated by the following statistical and practical considerations:

Weighted Correlation Test. There are two primary reasons:

  • ESS & Small-Sample Behavior: Learning decoupling weights acts similarly to sample selection, the better the decoupling, the smaller the effective sample size (ESS). In regimes with a limited ESS, complex nonparametric tests do not necessarily outperform simpler parametric ones:
    • Null Calibration: Complex nonparametric tests can suffer from poor asymptotic uniformity under the null hypothesis (H0​) in small samples, leading to inflated FDR.
    • Power: Simpler tests may achieve higher statistical power when the effective sample size is small.
  • Symmetry Breaking via Reweighting: As you mentioned, correlation tests typically lose power under strictly symmetric nonlinear dependencies. However, the reweighted distribution induced by the learned decoupling weights is highly unlikely to preserve the exact symmetry of the original distribution or accidentally exhibit symmetry after reweighting.

This analysis aligns closely with the Precision–Recall trade-off observed in our experiments: while the HSIC test can capture a broader class of dependencies (slightly improving Recall), its weaker Type I error control in small effective sample regimes leads to a drop in Precision.

Pairwise HSIC. In finite-sample estimation, achieving theoretically perfect decoupling weights is inherently difficult. Since the weights are learned in an unsupervised manner by optimizing empirical statistics, even if an empirical decoupling objective reaches its optimum, it remains a finite-sample estimator, meaning the corresponding population-level decoupling is rarely exact (even when the estimator is unbiased). Considering the trade-off between theoretical joint independence and the practical difficulty of gradient-based optimization, we ultimately chose the simpler and more stable Pairwise HSIC objective.

Q3. ESS and Runtime

In our experiments, to ensure an adequate sample size for statistical testing, the minimum ESS threshold is fixed at 100 (optimization stops early once this threshold is reached, with the maximum number of optimization steps capped at 200; please refer to the source code in the supplementary material for full implementation details). Following your suggestion, we measured the average runtime and final ESS over 20 MB identification runs on a single CPU core under the same setting as Section 4.1 Func:

dMBd_{\text{MB}}51020
time (s)5.1714.1236.68
ESS191.24 ± 64.27118.70 ± 15.42257.69 ± 18.10
AuthorsOpenReview ↗

Response to Reviewer oTF3 (Part 3/3)

Q4. Higher Dimensionality and Low Noise Regimes

Low Noise Regimes: We specifically investigated performance across varying noise levels in Section 4.2, where Figure 1 illustrates the complete performance curves spanning from near-zero to very high noise levels. As analyzed in Appendix C (Appendix D in the revised manuscript), learning effective decoupling weights currently relies on the presence of sufficient noise. Consequently, our method does not exhibit advantages in near-deterministic (extremely low-noise) regimes.

Higher Dimensionality: Following your suggestion, we conducted additional experiments under the same setting as Section 4.1 Func for higher dimensions d=25d=25 and d=30d=30 (note that PC with KCIT was omitted due to prohibitive runtime):

MethodSHD$_c$ ↓ (d=25d=25)F1 ↑ (d=25d=25)SHD$_c$ ↓ (d=30d=30)F1 ↑ (d=30d=30)
DCD21.8 ± 4.40.43 ± 0.1627.3 ± 5.10.38 ± 0.15
PC (FisherZ)33.4 ± 8.70.13 ± 0.1043.7 ± 6.30.08 ± 0.07
PC (E-FisherZ)27.1 ± 4.90.16 ± 0.1134.2 ± 4.50.12 ± 0.10
PC (E-KCIT)21.8 ± 2.50.17 ± 0.0928.1 ± 2.70.13 ± 0.07
GES31.0 ± 9.90.12 ± 0.1240.5 ± 5.50.06 ± 0.06
ICALiNGAM19.4 ± 3.00.30 ± 0.1623.7 ± 1.50.31 ± 0.07
DirectLiNGAM18.0 ± 2.30.39 ± 0.1323.1 ± 1.80.34 ± 0.09
Notears20.8 ± 1.70.23 ± 0.1025.7 ± 2.10.20 ± 0.11
GraNDAG53.1 ± 24.50.12 ± 0.0779.1 ± 30.00.09 ± 0.07
GOLEM17.9 ± 2.50.40 ± 0.1323.3 ± 2.20.32 ± 0.11

Although the absolute performance of our method decreases slightly as dimensionality grows, DCD achieves the highest F1 score across all baselines, and maintains a substantial advantage compared to PC and GES.


[1] Rajen Dinesh Shah and J. Peters. The hardness of conditional independence testing and the generalised covariance measure. The Annals of Statistics, 2018.

[2] Dimitris Margaritis, Sebastian Thrun. Bayesian Network Induction via Local Neighborhoods. NIPS, 1999.

[3] Ioannis Tsamardinos, Constantin F. Aliferis, and Alexander Statnikov. Algorithms for Large Scale Markov Blanket Discovery. FLAIRS, 2003.

[4] Jean-Philippe Pellet, André Elisseeff. Using Markov Blankets for Causal Structure Learning. JMLR, 2008.

AuthorsOpenReview ↗

Looking Forward to Your Feedback

Dear Reviewer oTF3,

As the discussion period is coming to a close, we are kindly wondering if any questions or concerns remain unresolved. We would be very happy to provide any further clarification.

Thank you again for your time and consideration.

Comments on the paper

1 comment

Dear Reviewers,

The author–reviewer discussion period is nearing its end, and the authors have responded to your reviews. Please engage with them as soon as possible, indicating whether their responses address your concerns and raising any remaining questions while there is still time for clarification.

Many thanks,
AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.