acceptodds
Under review as a conference paper at ICLR 2027

Local Constraint-Based Causal Discovery in the Presence of Cycles and Latent Variables

Abstract

Local causal discovery is of practical importance because many practical needs call only for identifying the causal structure surrounding a single target variable, making recovery of the entire global structure unnecessary. However, existing local methods typically assume acyclicity or the absence of latent variables. We study local causal discovery when cycles and latent variables are both present. We first characterize a local region sufficient for target-specific structure learning and establish which locally learned causal information is consistent with that obtained from global discovery. Building on these results, we propose a local constraint-based algorithm and prove that, under standard assumptions, it is sound and complete, recovering the same identifiable structure around the target as global causal discovery. Extensive experiments on random and real-world structures demonstrate the effectiveness and efficiency of our method.

Primary area:
causal reasoning
Keywords:
Local causal discovery, latent variables, directed cycles, $\sigma$-separation.

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_G48nOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates local causal discovery in the presence of both directed cycles and latent variables, and proposes a constraint-based method called LoCL. The authors first establish a graphical characterization of the Markov blanket and theoretical connections between local and global causal structures. Based on these results, they develop a local discovery algorithm that progressively expands the search region to recover the causal structure surrounding a target variable, with theoretical guarantees of soundness and completeness. Experiments on random cyclic graphs and the DREAM4 gene regulatory network demonstrate promising structural recovery accuracy and computational efficiency.

Critical strengths

1.The paper addresses local causal discovery in the simultaneous presence of directed cycles and latent variables, relaxing important structural assumptions made by most existing local methods. This makes the problem theoretically meaningful.

2.The paper establishes theoretical consistency between local and global causal structures and develops a local learning algorithm with theoretical guarantees. The ability to reduce computational costs while preserving identifiable causal information is a meaningful contribution.

Critical weaknesses

1.The paper does not include MMB-by-MMB (Xie et al., 2024)[1], a relevant local causal discovery method that considers latent variables. Including this baseline would provide a more comprehensive evaluation of LoCL against existing local causal discovery approaches.

2.Although the paper reports several quantitative metrics, it does not provide a direct visualization of the local causal structures recovered by LoCL on the experimental datasets. Such visualizations, together with the corresponding ground-truth structures, would help readers better understand the learned causal relationships and assess the reliability of the proposed method.

Reference:

[1] Local causal structure learning in the presence of latent variables. ICML24

Questions for authors

1.Why is MMB-by-MMB not included in the experimental comparisons? Could the authors clarify this omission and consider including it as an additional baseline?

2.Could the authors provide visual comparisons between the local causal structures recovered by LoCL and the corresponding ground-truth structures? This would help illustrate the method's ability to recover causal relationships beyond the reported quantitative metrics.

3.I noticed that your method can handle both circles and latent variables. Could the authors clarify the practical applications of handling circles and explain what real-world problems this capability can address?

5 comments
AuthorsOpenReview ↗

Response to Reviewer G48n - Part 1

We are grateful for your constructive comments and for recognizing the significance of handling directed cycles and latent variables, the theoretical consistency between local and global causal structures, and the ability to reduce computational costs while preserving identifiable causal information. We hope the responses below clarify the points raised.

Q1. The paper does not include MMB-by-MMB (Xie et al., 2024), a relevant local causal discovery method that considers latent variables. Including this baseline would provide a more comprehensive evaluation of LoCL against existing local causal discovery approaches. Why is MMB-by-MMB not included in the experimental comparisons? Could the authors clarify this omission and consider including it as an additional baseline?

A1. Thank you for your valuable suggestion. Following your suggestion, we have included MMB-by-MMB as an additional baseline in our experiments. We would also like to note that our original experiments included LatentLCD, a recent local method that accounts for latent variables.

The comparison results under the original settings are summarized in Table R10 below. LoCL consistently outperforms MMB-by-MMB across all settings, further demonstrating its effectiveness and efficiency in the presence of cycles and latent variables.

We have also conducted an additional comparison in an acyclic setting using the real-world ANDES network with 223 nodes, 8 latent variables, and a sample size of 1,000. As shown in the last two rows of Table R10, LoCL achieves slightly higher structural accuracy than MMB-by-MMB while incurring lower computational costs. We will incorporate these additional results into the revised manuscript.


Table R10. Additional baseline.

NetworknnMethodPrecision ↑Recall ↑F1 ↑Local SHD ↓Runtime (s)↓CI tests (×10³) ↓
d=501000LoCL0.85 ± 0.270.85 ± 0.270.85 ± 0.270.60 ± 1.100.03 ± 0.100.17 ± 0.15
MMB-by-MMB0.78 ± 0.340.78 ± 0.340.78 ± 0.340.87 ± 1.360.07 ± 0.040.38 ± 0.23
3000LoCL0.86 ± 0.290.86 ± 0.290.86 ± 0.290.55 ± 1.150.03 ± 0.090.29 ± 0.46
MMB-by-MMB0.80 ± 0.340.77 ± 0.350.78 ± 0.340.93 ± 1.390.13 ± 0.060.53 ± 0.40
5000LoCL0.90 ± 0.190.90 ± 0.190.90 ± 0.190.40 ± 0.750.03 ± 0.090.36 ± 0.49
MMB-by-MMB0.83 ± 0.280.83 ± 0.280.83 ± 0.280.67 ± 1.110.20 ± 0.120.77 ± 0.61
d=1001000LoCL0.82 ± 0.340.82 ± 0.340.82 ± 0.340.80 ± 1.640.03 ± 0.100.12 ± 0.16
MMB-by-MMB0.80 ± 0.450.80 ± 0.450.80 ± 0.451.20 ± 2.680.34 ± 0.410.45 ± 0.59
3000LoCL0.90 ± 0.310.90 ± 0.310.90 ± 0.310.60 ± 1.850.03 ± 0.090.13 ± 0.21
MMB-by-MMB0.65 ± 0.490.65 ± 0.490.65 ± 0.491.80 ± 2.680.43 ± 0.210.68 ± 0.73
5000LoCL0.95 ± 0.150.93 ± 0.180.93 ± 0.170.30 ± 0.730.04 ± 0.100.27 ± 0.39
MMB-by-MMB0.85 ± 0.340.85 ± 0.340.85 ± 0.340.60 ± 1.340.63 ± 0.310.71 ± 0.70
d=1501000LoCL0.88 ± 0.280.88 ± 0.280.88 ± 0.280.50 ± 1.100.04 ± 0.100.24 ± 0.40
MMB-by-MMB0.75 ± 0.430.75 ± 0.430.75 ± 0.431.00 ± 1.731.65 ± 1.640.82 ± 0.54
3000LoCL0.95 ± 0.170.95 ± 0.170.95 ± 0.170.20 ± 0.700.03 ± 0.100.15 ± 0.19
MMB-by-MMB0.89 ± 0.250.89 ± 0.250.89 ± 0.250.44 ± 1.012.29 ± 2.660.89 ± 0.82
5000LoCL0.96 ± 0.090.96 ± 0.090.96 ± 0.090.15 ± 0.370.04 ± 0.100.50 ± 0.89
MMB-by-MMB0.86 ± 0.250.86 ± 0.250.86 ± 0.250.56 ± 1.013.07 ± 3.461.12 ± 0.98
DREAM41000LoCL0.59 ± 0.310.47 ± 0.110.50 ± 0.162.70 ± 1.630.11 ± 0.144.21 ± 3.40
MMB-by-MMB0.30 ± 0.230.30 ± 0.230.30 ± 0.232.80 ± 0.920.69 ± 0.385.08 ± 4.85
ANDES1000LoCL0.90 ± 0.230.87 ± 0.240.88 ± 0.230.80 ± 1.510.31 ± 0.290.12 ± 0.08
MMB-by-MMB0.88 ± 0.180.85 ± 0.210.86 ± 0.190.90 ± 1.254.81 ± 2.960.53 ± 0.77
AuthorsOpenReview ↗

Response to Reviewer G48n - Part 2

Q2. Although the paper reports several quantitative metrics, it does not provide a direct visualization of the local causal structures recovered by LoCL on the experimental datasets. Such visualizations, together with the corresponding ground-truth structures, would help readers better understand the learned causal relationships and assess the reliability of the proposed method. Could the authors provide visual comparisons between the local causal structures recovered by LoCL and the corresponding ground-truth structures? This would help illustrate the method's ability to recover causal relationships beyond the reported quantitative metrics.

A2. Thank you for your helpful suggestion. To provide a more intuitive illustration of LoCL's structural recovery performance, we present two examples from the DREAM4 dataset, comparing the ground-truth σ-PAGs with the corresponding local structures recovered by LoCL around targets G88 and G70. The corresponding target-adjacent edges are aligned below for comparison.

Ground TruthLoCL
G85 ∘→\circ \rightarrow G88G85 ↔\leftrightarrow G88
G62 ↔\leftrightarrow G88No edge
G44 ↔\leftrightarrow G88G44 ↔\leftrightarrow G88
G37 ↔\leftrightarrow G88No edge
G10 ↔\leftrightarrow G88G10 ↔\leftrightarrow G88
Ground TruthLoCL
G37 →\rightarrow G70G37 →\rightarrow G70
G28 ↔\leftrightarrow G70G28 ↔\leftrightarrow G70

As shown above, LoCL perfectly recovers the local structure around G70. For G88, LoCL correctly recovers two bidirected edges but misses the edges between G88 and G62 and between G88 and G37. It also incorrectly orients the circle endpoint at G85 as an arrowhead. We will include these visualizations in the revised manuscript.


Q3. I noticed that your method can handle both circles and latent variables. Could the authors clarify the practical applications of handling circles and explain what real-world problems this capability can address?

A3. Thank you for this insightful question. Handling causal cycles is particularly important for real-world systems involving feedback mechanisms. For example, Mooij and Heskes (2013) demonstrated the practical relevance of cyclic causal discovery by reconstructing a cellular signaling network from real-world flow cytometry data, revealing evidence of feedback loops. Similarly, feedback mechanisms are widely recognized in gene regulatory networks (Rohbeck et al., 2024) and neural systems (Sanchez-Romero et al., 2019).

In such systems, conventional acyclicity assumptions may exclude important causal relationships. Handling cycles therefore extends the applicability of causal discovery to complex systems with feedback (Forré and Mooij, 2018; Mooij and Claassen, 2020), enabling a more faithful representation of causal relationships and providing a foundation for subsequent causal reasoning and intervention analysis (Forré and Mooij, 2020). We will include this discussion in the revised manuscript.

References:

  • Joris M. Mooij and Tom Heskes. Cyclic causal discovery from continuous equilibrium data. UAI, 2013.
  • Martin Rohbeck, Brian Clarke, Katharina Mikulik, Alexandra Pettet, Oliver Stegle, and Kai Ueltzhöffer. Bicycle: Intervention-based causal discovery with cycles. CLeaR, 2024.
  • Ruben Sanchez-Romero, Joseph D. Ramsey, Kun Zhang, Madelyn R. K. Glymour, Biwei Huang, and Clark Glymour. Estimating feedforward and feedback effective connections from fMRI time series: Assessments of statistical methods. Network Neuroscience, 2019.
  • Patrick Forré and Joris M. Mooij. Constraint-based causal discovery for non-linear structural causal models with cycles and latent confounders. UAI, 2018.
  • Joris M. Mooij and Tom Claassen. Constraint-based causal discovery using partial ancestral graphs in the presence of cycles. UAI, 2020.
  • Patrick Forré and Joris M. Mooij. Causal calculus in the presence of cycles, latent confounders and selection bias. UAI, 2020.

Dear Reviewer G48n,

We thank you once again for your time and efforts. Since the discussion period will end in 2 days, we were wondering whether our response has addressed your concerns. If you have any additional comments, please let us know, we would be happy to address them.

Thanks a lot,

The authors

Reviewer_Wm1cOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a local constraint-based causal discovery algorithm, named LoCL, designed to identify causal relationships surrounding a target variable in systems containing both directed cycles and latent confounders. The approach builds on σ-separation, σ-MAGs, and MB discovery. The authors establish theoretical relationships between locally recovered causal information and global causal structures, and claim soundness and completeness relative to global σ-FCI under σ-Markov and σ-faithfulness assumptions. Experimental evaluations on synthetic cyclic graphs and a DREAM4 gene regulatory network suggest improvements in computational efficiency and target-specific structural accuracy compared with several global and local causal discovery methods.

Critical strengths

The paper addresses an interesting problem by considering both cycles and latent confounding in local causal discovery. The theoretical analysis of local–global consistency is a positive aspect, and the proposed method offers a reasonable way to reduce computational costs through MB localization. The obtained efficiency improvements, particularly in the number of conditional independence tests, are encouraging.

Critical weaknesses

The paper suffer from several important limitations that weaken its contribution.

  • The novelty is very limited, as LoCL mainly combines existing MB discovery techniques, FCI, and established orientation rules. The theoretical guarantees rely on strong assumptions, particularly σ-faithfulness and accurate conditional independence information, without investigating how violations of these assumptions affect the results. The complexity analysis is also incomplete, since the claimed efficiency depends heavily on the size of the MBs, which may become very large in complex causal structures.

  • The experimental evaluation represents the main drawback of the paper. All experiments rely on linear Gaussian models, with relatively small graphs and limited latent confounding, making it difficult to assess the method's generalizability.

  • Several baselines were not designed to handle cycles and latent variables simultaneously, raising concerns about the fairness of the comparisons. The DREAM4 evaluation is also unconvincing, with an F1 score of only 0.50 and recall of 0.47.

  • LoCL does not outperform L-MARVEL in terms of Local-SHD. The absence of nonlinear experiments, and ablation studies weakens the empirical evidence.

  • The last thing is that the authors did not investigate how errors in MB estimation propagate through the algorithm, which is critical for practical applications.

Questions for authors

No questions to be reported.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Wm1c - Part 1

We appreciate the time and care you devoted to our paper, as well as your recognition of the interesting problem, the theoretical analysis of local–global consistency, and the encouraging efficiency improvements, particularly in the number of CI tests. We hope the following response properly addresses your concerns.

Q1. The novelty is very limited, as LoCL mainly combines existing MB discovery techniques, FCI, and established orientation rules. The theoretical guarantees rely on strong assumptions, particularly σ-faithfulness and accurate conditional independence information, without investigating how violations of these assumptions affect the results. The complexity analysis is also incomplete, since the claimed efficiency depends heavily on the size of the MBs, which may become very large in complex causal structures.

A1. Thank you for your comments. We would like to clarify the following three points.

1. Technical contribution. We would like to clarify our contributions from two aspects.

  • Methodologically, we propose LoCL, a local method that jointly addresses cycles and latent confounding, extending local causal discovery to a more general setting.
  • Theoretically, we establish local-to-global consistency theoretical results that characterize which adjacencies, collider paths, and endpoint marks can be preserved through local learning (Theorems 1–3). Building on these results, Theorem 4 guarantees that LoCL recovers the same target adjacencies and invariant endpoint marks as global discovery.

2. Assumptions. We would like to clarify that our method relaxes two important assumptions commonly made by existing local causal discovery methods: acyclicity and the absence of latent confounding. We would like to clarify three additional points.

  • σ-Faithfulness and oracle CI information are also required for the theoretical guarantees of global σ-FCI. The theoretical guarantees of classical causal discovery methods, such as PC, FCI, and their extensions, typically rely on faithfulness and oracle CI information. These methods have been used in a range of fields (Spirtes and Zhang, 2016).
  • In our experiments, we perform CI tests on finite samples rather than relying on oracle CI information. Moreover, LoCL shows improvements in structural accuracy as the sample size increases in our experiments, providing empirical support for its effectiveness with sample-based CI tests.
  • Our experiments do not explicitly enforce σ-faithfulness, yet the results demonstrate LoCL's empirical effectiveness. Moreover, local learning approaches such as ours significantly reduce unnecessary CI tests, compared with global approaches. This alleviates the instability caused by excessive CI testing in high-dimensional settings and may help mitigate the effects of minor faithfulness violations (Isozaki, 2014). The additional experiment on σ-faithfulness violations in A4 shows a similar trend.

3. Efficiency and MB size. We would like to clarify that small MBs have long been a key motivation for local causal discovery, as they are considered plausible in many real-world problems (Margaritis and Thrun, 1999; Tsamardinos et al., 2003; Aliferis et al., 2003, 2010; Peña et al., 2007). Moreover, in highly complex causal systems where larger MBs may arise, their impact is limited to reducing LoCL's computational advantage over global learning methods, without compromising its soundness and completeness.

We will include these discussions in the revised manuscript.

References:

  • Peter Spirtes and Kun Zhang. Causal discovery and inference: concepts and recent methodological advances. Applied Informatics, 2016.
  • Takashi Isozaki. A robust causal discovery algorithm against faithfulness violation. Information and Media Technologies, 2014.
  • Margaritis and Thrun. Bayesian Network Induction via Local Neighborhoods. NeurIPS, 1999.
  • Tsamardinos et al. Time and Sample Efficient Discovery of Markov Blankets and Direct Causal Relations. KDD, 2003.
  • Aliferis et al. HITON: A Novel Markov Blanket Algorithm for Optimal Variable Selection. AMIA Annual Symposium, 2003.
  • Aliferis et al. Local Causal and Markov Blanket Induction for Causal Discovery and Feature Selection for Classification. Part I: Algorithms and Empirical Evaluation. JMLR, 2010.
  • Peña et al. Towards Scalable and Data Efficient Learning of Markov Boundaries. IJAR, 2007.
AuthorsOpenReview ↗

Response to Reviewer Wm1c - Part 2

Q2. The experimental evaluation represents the main drawback of the paper. All experiments rely on linear Gaussian models, with relatively small graphs and limited latent confounding, making it difficult to assess the method's generalizability.

A2. Thank you for the helpful suggestion. We have expanded the experiments in four directions.

1. Linear non-Gaussian models. We evaluated all methods on 100-node linear cyclic SCMs with Uniform and Laplace noise. The results are shown in Tables R1 and R2 in our response to Reviewer mV61. LoCL achieves higher structural accuracy at lower computational cost across both noise distributions and all sample sizes. These results support its accuracy and efficiency beyond Gaussian noise.

2. Nonlinear feedback models. We evaluated all methods on 100-node nonlinear cyclic SCMs with Gaussian and Uniform noise. The results are shown in Tables R3 and R4 in our response to Reviewer mV61. LoCL also achieves the highest F1 and lowest Local-SHD across all settings, while maintaining the shortest runtime. These results support LoCL's empirical effectiveness beyond linear models.

3. Larger structures. We added experiments on 500-node graphs with 25 latent variables and eight enforced directed cycles of length five, with all other settings consistent with those in our manuscript. The results are shown in Table R5 below. LoCL maintains its advantages in structural accuracy and computational efficiency on these larger networks. Its computational advantage also shows a tendency to increase with graph size. For example, at n=5000n=5000, LoCL is approximately 20× faster than σ-FCI on the 100-node graphs (0.04 s versus 0.80 s), compared with 31.2× on the 500-node graphs evaluated here (0.50 s versus 15.62 s).

4. Latent confounding. We varied the number of latent variables from 5 to 20 on 100-node graphs. The results are shown in Table R6 below. As expected, the structural accuracy of all methods decreases as the number of latent variables increases from 5 to 20. Nevertheless, even with 20 latent variables, LoCL maintains its advantages in structural accuracy over existing local methods and in computational efficiency over global methods.

Together, these additional experiments demonstrate LoCL's effectiveness and efficiency beyond linear Gaussian models, including larger networks and settings with more latent variables. We will include these results in the revised manuscript.


Table R5. Larger 500-node random structures.

MethodPrecision ↑Recall ↑F1 ↑Local SHD ↓Runtime (s) ↓CI tests (×10³) ↓
Sample size n=1000n=1000
σ-FCI0.88 ± 0.110.85 ± 0.140.86 ± 0.091.00 ± 0.739.84 ± 0.73208.28 ± 26.28
L-MARVEL0.88 ± 0.250.88 ± 0.250.88 ± 0.250.75 ± 1.336.04 ± 5.07195.56 ± 78.48
MMB-by-MMB0.72 ± 0.270.69 ± 0.250.70 ± 0.251.67 ± 1.22245.26 ± 289.091.16 ± 0.85
LatentLCD0.56 ± 0.270.60 ± 0.220.58 ± 0.253.10 ± 2.4512.40 ± 10.92150.93 ± 135.94
MB-by-MB0.28 ± 0.260.46 ± 0.330.34 ± 0.287.95 ± 4.529.74 ± 6.7733.00 ± 35.98
CMB0.50 ± 0.290.50 ± 0.280.50 ± 0.283.25 ± 2.1776.15 ± 127.79443.63 ± 904.09
LoCL (ours)0.92 ± 0.140.92 ± 0.140.92 ± 0.130.60 ± 0.990.47 ± 0.020.13 ± 0.16
Sample size n=3000n=3000
σ-FCI0.85 ± 0.240.86 ± 0.220.85 ± 0.221.20 ± 1.7412.72 ± 2.31277.14 ± 57.27
L-MARVEL0.90 ± 0.130.94 ± 0.100.91 ± 0.100.65 ± 0.81117.16 ± 177.431105.54 ± 1556.17
MMB-by-MMB0.87 ± 0.220.73 ± 0.190.79 ± 0.181.60 ± 1.14153.80 ± 151.411.17 ± 1.19
LatentLCD0.61 ± 0.230.71 ± 0.210.65 ± 0.212.80 ± 1.8221.78 ± 16.95327.95 ± 275.92
MB-by-MB0.26 ± 0.240.46 ± 0.300.32 ± 0.259.30 ± 5.0871.78 ± 62.38110.52 ± 134.25
CMB0.51 ± 0.240.57 ± 0.240.53 ± 0.233.12 ± 1.27491.00 ± 594.47565.25 ± 702.54
LoCL (ours)0.97 ± 0.060.96 ± 0.090.97 ± 0.070.25 ± 0.550.48 ± 0.030.25 ± 0.51
Sample size n=5000n=5000
σ-FCI0.81 ± 0.240.81 ± 0.210.81 ± 0.221.20 ± 1.4015.62 ± 6.29316.14 ± 66.74
L-MARVEL0.88 ± 0.240.91 ± 0.220.89 ± 0.230.75 ± 1.45171.31 ± 380.231303.75 ± 2553.83
MMB-by-MMB0.83 ± 0.240.75 ± 0.220.78 ± 0.211.50 ± 1.29310.76 ± 426.821.98 ± 3.19
LatentLCD0.60 ± 0.300.67 ± 0.280.63 ± 0.293.00 ± 2.7026.96 ± 17.69462.69 ± 318.01
MB-by-MB0.27 ± 0.260.45 ± 0.310.32 ± 0.279.35 ± 5.75150.21 ± 145.40144.91 ± 214.42
CMB0.48 ± 0.270.49 ± 0.270.49 ± 0.272.93 ± 1.77379.43 ± 552.06291.65 ± 483.25
LoCL (ours)0.98 ± 0.050.97 ± 0.090.97 ± 0.070.20 ± 0.520.50 ± 0.050.28 ± 0.63
AuthorsOpenReview ↗

Response to Reviewer Wm1c - Part 3

Table R6. Varying the number of latent variables.

MethodPrecision ↑Recall ↑F1 ↑Local SHD ↓Runtime (s) ↓CI tests (×10³) ↓
Latent variables ∣L∣=5\lvert L\rvert=5
σ-FCI0.90 ± 0.170.88 ± 0.190.88 ± 0.180.50 ± 0.760.40 ± 0.0612.75 ± 2.54
L-MARVEL0.90 ± 0.260.93 ± 0.230.91 ± 0.240.60 ± 1.700.25 ± 0.328.32 ± 5.83
MMB-by-MMB0.88 ± 0.260.88 ± 0.260.88 ± 0.260.50 ± 1.050.84 ± 0.620.47 ± 0.39
LatentLCD0.81 ± 0.270.84 ± 0.260.82 ± 0.260.95 ± 1.570.57 ± 0.398.03 ± 5.30
MB-by-MB0.31 ± 0.330.49 ± 0.360.36 ± 0.336.75 ± 4.200.84 ± 0.666.69 ± 6.44
CMB0.40 ± 0.400.41 ± 0.400.41 ± 0.402.55 ± 1.792.42 ± 4.6624.13 ± 53.14
LoCL (ours)0.95 ± 0.130.94 ± 0.140.94 ± 0.130.40 ± 0.990.03 ± 0.090.18 ± 0.29
Latent variables ∣L∣=10\lvert L\rvert=10
σ-FCI0.84 ± 0.270.84 ± 0.260.84 ± 0.260.80 ± 1.200.39 ± 0.0611.78 ± 1.81
L-MARVEL0.90 ± 0.190.91 ± 0.190.90 ± 0.190.65 ± 1.310.17 ± 0.076.23 ± 1.09
MMB-by-MMB0.73 ± 0.330.71 ± 0.330.72 ± 0.331.55 ± 1.970.87 ± 0.570.49 ± 0.36
LatentLCD0.65 ± 0.300.69 ± 0.250.66 ± 0.272.10 ± 1.940.65 ± 0.3910.70 ± 10.27
MB-by-MB0.25 ± 0.270.42 ± 0.380.31 ± 0.307.50 ± 4.650.74 ± 0.435.95 ± 3.81
CMB0.41 ± 0.390.40 ± 0.380.40 ± 0.382.95 ± 1.901.84 ± 2.8918.81 ± 33.31
LoCL (ours)0.91 ± 0.200.94 ± 0.180.92 ± 0.190.45 ± 1.100.03 ± 0.100.14 ± 0.18
Latent variables ∣L∣=20\lvert L\rvert=20
σ-FCI0.83 ± 0.270.79 ± 0.260.80 ± 0.261.00 ± 1.210.32 ± 0.059.31 ± 1.99
L-MARVEL0.85 ± 0.250.86 ± 0.240.85 ± 0.240.80 ± 1.240.17 ± 0.115.52 ± 1.84
MMB-by-MMB0.68 ± 0.290.71 ± 0.300.69 ± 0.291.60 ± 1.680.49 ± 0.290.41 ± 0.30
LatentLCD0.64 ± 0.300.66 ± 0.280.65 ± 0.292.05 ± 1.760.46 ± 0.374.94 ± 3.55
MB-by-MB0.18 ± 0.250.36 ± 0.390.23 ± 0.289.25 ± 5.161.13 ± 1.0910.14 ± 11.43
CMB0.32 ± 0.390.35 ± 0.400.33 ± 0.393.50 ± 2.191.63 ± 2.4916.27 ± 28.95
LoCL (ours)0.86 ± 0.230.87 ± 0.210.86 ± 0.220.70 ± 1.080.03 ± 0.090.11 ± 0.11

Q3. Several baselines were not designed to handle cycles and latent variables simultaneously, raising concerns about the fairness of the comparisons. The DREAM4 evaluation is also unconvincing, with an F1 score of only 0.50 and recall of 0.47.

A3. We would like to clarify the following two points:

  • Our primary motivation is the lack of local causal discovery methods that jointly handle directed cycles and latent variables. Furthermore, our theoretical guarantees also hold in acyclic or causal sufficiency settings. To further demonstrate the applicability of LoCL, we conducted an additional experiment on the acyclic ANDES network from the bnlearn repository, which contains 223 nodes and 338 edges. We randomly selected 8 variables as latent and set the sample size to 1,000. The results are presented in Table R7 below. LoCL achieves the best performance across all evaluated metrics, further supporting its effectiveness and efficiency in acyclic settings.

  • Recovery accuracy remains limited across all compared methods on DREAM4. The relatively modest absolute accuracy may reflect the complexity of this biologically grounded network (Schaffter et al., 2011). Nevertheless, LoCL performs best on almost all metrics. Improving structural recovery accuracy in such challenging settings remains an important direction for future work (Marbach et al., 2012).


Table R7. ANDES acyclic Network.

MethodPrecision ↑Recall ↑F1 ↑Local SHD ↓Runtime (s) ↓CI tests (×10³) ↓
FCI0.86 ± 0.240.85 ± 0.230.85 ± 0.231.05 ± 1.473.38 ± 0.36108.21 ± 11.87
L-MARVEL0.89 ± 0.190.85 ± 0.240.86 ± 0.221.00 ± 1.724.22 ± 3.8484.25 ± 61.93
MMB-by-MMB0.88 ± 0.180.85 ± 0.210.86 ± 0.190.90 ± 1.254.81 ± 2.960.53 ± 0.77
LatentLCD0.73 ± 0.210.74 ± 0.180.73 ± 0.192.10 ± 2.003.67 ± 5.66133.39 ± 265.09
MB-by-MB0.28 ± 0.110.67 ± 0.150.38 ± 0.1112.00 ± 5.1811.80 ± 15.4388.55 ± 133.11
CMB0.66 ± 0.200.65 ± 0.210.65 ± 0.212.40 ± 1.73280.48 ± 742.352728.87 ± 7519.50
LoCL (ours)0.90 ± 0.230.87 ± 0.240.88 ± 0.230.80 ± 1.510.31 ± 0.290.12 ± 0.08

References:

  • Schaffter et al. GeneNetWeaver: in silico benchmark generation and performance profiling of network inference methods. Bioinformatics, 2011.
  • Marbach et al. Wisdom of crowds for robust gene network inference. Nature Methods, 2012.
Reviewer_mV61OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper focuses on local constraint-based causal discovery with cycles and latent variables. Unlike previous methods that address either global CD with cycles and latent variables or local CD with latent variables or cycles, this work claims to be the first local CD method with both cycles and latent variables.

Critical strengths
  1. The presentation is clear overall, and the related work is extensively studied and compared.
  2. The core of the paper is to prove the local–global correspondence within a single MB region, discuss its limitations, and propose LoCL.
  3. The empirical results on synthetic data show clear efficiency and effectiveness gains from LoCL compared to the baselines.
Critical weaknesses
  1. The synthetic experiments use linear Gaussian cyclic models, which can violate the required σ-faithfulness assumption. Please check whether σ-faithfulness holds in your experiments.
  2. Since the theory is broader, please include synthetic experiments on linear non-Gaussian models and nonlinear feedback models.
Questions for authors
  1. In Sec. 4.2, you discussed the limitations of using a single Markov blanket region. How do these limitations affect the experimental results?
11 comments
AuthorsOpenReview ↗

Response to Reviewer mV61

We thank you for your thoughtful review and for highlighting the clear presentation, extensive study and comparison of related work, local–global correspondence within a single MB region, and clear efficiency and effectiveness gains of LoCL. We address your comments point by point below.

Q1. The synthetic experiments use linear Gaussian cyclic models, which can violate the required σ-faithfulness assumption. Please check whether σ-faithfulness holds in your experiments.

A1. Thank you for raising this important point. Our theoretical guarantee relies on σ-faithfulness, as does that of global σ-FCI. Our experiments do not explicitly enforce this assumption, but the results demonstrate LoCL’s empirical effectiveness. We would like to clarify three additional points.

  • First, faithfulness is a standard assumption underlying classical constraint-based causal discovery methods, such as PC, FCI, and their many extensions, which have been widely applied across various fields (Spirtes and Zhang, 2016).
  • Second, local learning approaches such as LoCL significantly reduce the number of unnecessary CI tests compared with global methods. This may alleviate the instability of extensive CI testing in high-dimensional settings and potentially reduce sensitivity to minor faithfulness violations (Isozaki, 2014). Our additional experiments on σ-faithfulness violations (Table R8, response to Reviewer Wm1c) further support this potential advantage.
  • Finally, incorporating elements of Greedy Equivalence Search (GES) (Chickering, 2002), a representative score-based method, may improve robustness to faithfulness violations (Zhalama et al., 2017). We plan to explore this direction in future work.

We will include this discussion in the revised manuscript.

References:

  • Peter Spirtes and Kun Zhang. Causal discovery and inference: concepts and recent methodological advances. Applied Informatics, 2016.
  • Takashi Isozaki. A robust causal discovery algorithm against faithfulness violation. Information and Media Technologies, 2014.
  • David Maxwell Chickering. Optimal structure identification with greedy search. Journal of Machine Learning Research, 2002.
  • Zhalama, Jiji Zhang, and Wolfgang Mayer. Weakening faithfulness: some heuristic causal discovery algorithms. International Journal of Data Science and Analytics, 2017.

Q2. Since the theory is broader, please include synthetic experiments on linear non-Gaussian models and nonlinear feedback models.

A2. Thank you for the helpful suggestion. We have added experiments on linear non-Gaussian and nonlinear feedback models under the random-structure setting with 100 nodes and sample sizes of 1,000, 3,000, and 5,000. Other settings remain the same as those in our manuscript.

(1) Linear non-Gaussian models. We evaluated linear cyclic SCMs with Uniform(−3,3)\mathrm{Uniform}(-\sqrt{3},\sqrt{3}) or Laplace(0,1/2)\mathrm{Laplace}(0,1/\sqrt{2}) noise. Tables R1 and R2 report the results.

(2) Nonlinear feedback models. Following Garcez et al. (2026), we evaluated nonlinear feedback SCMs of the form V=tanh⁡(WV)+UV=\tanh(WV)+U with standard Gaussian or Uniform(−3,3)\mathrm{Uniform}(-\sqrt{3},\sqrt{3}) noise. Tables R3 and R4 report the results.

As shown in Tables R1–R4, LoCL outperforms the other methods on all evaluation metrics across nearly all combinations of the four settings and three sample sizes. These results further support LoCL’s empirical effectiveness and efficiency beyond the linear Gaussian setting. Note that all methods use partial-correlation-based conditional independence tests in the nonlinear settings. Thus, like σ-FCI and L-MARVEL, LoCL is less effective in the nonlinear settings than in the linear settings. Adopting more suitable tests for nonlinear models could improve performance. We will add these results to the revised version.

References:

  • Ana Letícia Garcez Vicente, Gijs van Seeventer, and Saber Salehkaleybar. Data-Driven Covariate Selection for Nonparametric and Cycle-Agnostic Causal Effect Estimation. NeurIPS, 2026.

Q3. In Sec. 4.2, you discussed the limitations of using a single Markov blanket region. How do these limitations affect the experimental results?

A3. Thank you for this question. As discussed in Sec. 4.2, restricting learning to a single MB may leave some globally identifiable endpoints unoriented. LoCL addresses this limitation by exploring additional MBs to resolve these orientations, as illustrated in Example 5.

In practice, this expansion incurs additional computational cost when the required information is unavailable within the initial MB. However, its overall cost remains substantially lower than that of global learning, as our experiments show, while LoCL retains the same identifiability guarantees.

AuthorsOpenReview ↗

Tables R1–R2: Linear Non-Gaussian Models

Table R1. Linear Uniform models.

MethodPrecision ↑Recall ↑F1 ↑Local SHD ↓Runtime (s) ↓CI tests (×10³) ↓
Sample size n=1000n=1000
σ-FCI0.82 ± 0.290.79 ± 0.270.79 ± 0.271.25 ± 1.870.45 ± 0.0713.55 ± 2.54
L-MARVEL0.74 ± 0.380.74 ± 0.380.74 ± 0.381.25 ± 1.970.24 ± 0.157.80 ± 2.58
MMB-by-MMB0.75 ± 0.420.75 ± 0.420.75 ± 0.421.20 ± 2.150.26 ± 0.090.40 ± 0.28
LatentLCD0.72 ± 0.290.74 ± 0.270.73 ± 0.281.35 ± 1.740.69 ± 0.4010.04 ± 6.63
MB-by-MB0.25 ± 0.290.40 ± 0.340.29 ± 0.297.10 ± 4.401.08 ± 0.969.32 ± 10.14
CMB0.48 ± 0.360.50 ± 0.380.49 ± 0.372.20 ± 1.636.53 ± 9.4559.22 ± 94.45
LoCL (ours)0.86 ± 0.310.86 ± 0.310.86 ± 0.310.65 ± 1.560.03 ± 0.100.15 ± 0.23
Sample size n=3000n=3000
σ-FCI0.84 ± 0.270.80 ± 0.270.81 ± 0.260.90 ± 1.260.60 ± 0.1417.98 ± 3.90
L-MARVEL0.85 ± 0.320.86 ± 0.310.86 ± 0.320.75 ± 1.700.33 ± 0.268.34 ± 3.43
MMB-by-MMB0.79 ± 0.310.82 ± 0.310.81 ± 0.311.10 ± 1.850.93 ± 0.850.85 ± 0.76
LatentLCD0.80 ± 0.260.83 ± 0.230.81 ± 0.241.00 ± 1.450.96 ± 0.6519.08 ± 13.53
MB-by-MB0.23 ± 0.270.39 ± 0.350.27 ± 0.297.95 ± 5.213.19 ± 2.9716.50 ± 17.48
CMB0.50 ± 0.400.50 ± 0.400.50 ± 0.402.10 ± 1.7612.48 ± 15.9669.74 ± 93.70
LoCL (ours)0.88 ± 0.270.89 ± 0.260.88 ± 0.260.65 ± 1.560.03 ± 0.090.13 ± 0.21
Sample size n=5000n=5000
σ-FCI0.95 ± 0.130.93 ± 0.160.93 ± 0.140.30 ± 0.640.84 ± 0.4326.59 ± 12.34
L-MARVEL0.81 ± 0.330.83 ± 0.330.82 ± 0.320.90 ± 1.480.58 ± 0.479.90 ± 4.57
MMB-by-MMB0.72 ± 0.360.68 ± 0.350.69 ± 0.351.30 ± 1.421.32 ± 1.070.80 ± 0.53
LatentLCD0.77 ± 0.280.80 ± 0.230.78 ± 0.251.40 ± 1.771.27 ± 0.7728.69 ± 20.53
MB-by-MB0.23 ± 0.260.41 ± 0.330.28 ± 0.278.15 ± 4.768.78 ± 6.3727.19 ± 21.87
CMB0.48 ± 0.380.48 ± 0.380.48 ± 0.382.30 ± 1.7920.60 ± 23.8579.34 ± 103.73
LoCL (ours)0.99 ± 0.050.96 ± 0.120.97 ± 0.090.15 ± 0.480.03 ± 0.090.23 ± 0.37

Table R2. Linear Laplace models.

MethodPrecision ↑Recall ↑F1 ↑Local SHD ↓Runtime (s) ↓CI tests (×10³) ↓
Sample size n=1000n=1000
σ-FCI0.76 ± 0.280.75 ± 0.260.75 ± 0.271.40 ± 1.880.50 ± 0.0713.43 ± 2.55
L-MARVEL0.76 ± 0.370.79 ± 0.370.77 ± 0.371.15 ± 1.820.25 ± 0.178.07 ± 2.96
MMB-by-MMB0.75 ± 0.420.75 ± 0.420.75 ± 0.421.20 ± 2.150.29 ± 0.100.40 ± 0.28
LatentLCD0.65 ± 0.290.70 ± 0.250.67 ± 0.271.80 ± 1.720.70 ± 0.4110.30 ± 6.61
MB-by-MB0.24 ± 0.270.40 ± 0.340.29 ± 0.287.20 ± 4.791.38 ± 2.0612.90 ± 23.31
CMB0.44 ± 0.380.48 ± 0.390.46 ± 0.382.50 ± 1.755.61 ± 8.8354.55 ± 96.56
LoCL (ours)0.86 ± 0.300.86 ± 0.290.86 ± 0.300.75 ± 1.700.03 ± 0.090.16 ± 0.24
Sample size n=3000n=3000
σ-FCI0.74 ± 0.300.74 ± 0.270.73 ± 0.281.35 ± 1.490.59 ± 0.1518.31 ± 3.91
L-MARVEL0.85 ± 0.340.85 ± 0.330.85 ± 0.330.80 ± 1.810.29 ± 0.177.57 ± 2.06
MMB-by-MMB0.68 ± 0.430.68 ± 0.430.68 ± 0.431.50 ± 2.120.76 ± 0.450.77 ± 0.71
LatentLCD0.73 ± 0.320.75 ± 0.300.74 ± 0.311.40 ± 1.850.85 ± 0.6316.06 ± 13.61
MB-by-MB0.22 ± 0.270.39 ± 0.350.27 ± 0.298.55 ± 5.573.40 ± 3.4417.72 ± 19.86
CMB0.46 ± 0.400.48 ± 0.410.47 ± 0.402.50 ± 2.0113.30 ± 13.9970.60 ± 79.75
LoCL (ours)0.91 ± 0.240.91 ± 0.240.91 ± 0.240.45 ± 1.360.03 ± 0.090.13 ± 0.21
Sample size n=5000n=5000
σ-FCI0.94 ± 0.130.91 ± 0.160.92 ± 0.150.35 ± 0.650.81 ± 0.3824.95 ± 12.20
L-MARVEL0.84 ± 0.310.85 ± 0.320.84 ± 0.310.80 ± 1.330.68 ± 0.7510.85 ± 7.30
MMB-by-MMB0.75 ± 0.370.70 ± 0.370.72 ± 0.361.20 ± 1.481.27 ± 1.010.75 ± 0.50
LatentLCD0.82 ± 0.280.81 ± 0.260.81 ± 0.271.05 ± 1.601.20 ± 0.8427.82 ± 21.00
MB-by-MB0.22 ± 0.260.40 ± 0.340.27 ± 0.278.30 ± 4.748.36 ± 6.0726.99 ± 21.24
CMB0.48 ± 0.380.48 ± 0.380.48 ± 0.382.50 ± 2.1118.80 ± 24.3978.37 ± 115.87
LoCL (ours)0.99 ± 0.050.96 ± 0.120.97 ± 0.090.15 ± 0.480.03 ± 0.090.21 ± 0.31
AuthorsOpenReview ↗

Tables R3–R4: Nonlinear Feedback Models

Table R3. Nonlinear Gaussian models.

MethodPrecision ↑Recall ↑F1 ↑Local SHD ↓Runtime (s) ↓CI tests (×10³) ↓
Sample size n=1000n=1000
σ-FCI0.54 ± 0.260.54 ± 0.230.53 ± 0.242.15 ± 1.190.34 ± 0.067.06 ± 1.04
L-MARVEL0.36 ± 0.360.39 ± 0.360.37 ± 0.362.95 ± 1.770.04 ± 0.014.67 ± 0.04
MMB-by-MMB0.36 ± 0.340.36 ± 0.340.36 ± 0.342.55 ± 1.361.19 ± 0.350.78 ± 0.26
LatentLCD0.46 ± 0.230.49 ± 0.190.46 ± 0.192.85 ± 1.420.48 ± 0.353.35 ± 2.63
MB-by-MB0.26 ± 0.270.33 ± 0.290.28 ± 0.274.10 ± 1.890.82 ± 0.486.73 ± 4.97
CMB0.29 ± 0.330.29 ± 0.330.29 ± 0.333.45 ± 1.881.74 ± 1.4115.54 ± 13.56
LoCL (ours)0.55 ± 0.260.55 ± 0.260.55 ± 0.261.80 ± 1.030.03 ± 0.090.04 ± 0.03
Sample size n=3000n=3000
σ-FCI0.57 ± 0.280.61 ± 0.280.59 ± 0.282.05 ± 1.560.46 ± 0.119.76 ± 2.99
L-MARVEL0.62 ± 0.390.64 ± 0.380.63 ± 0.391.95 ± 2.290.06 ± 0.014.81 ± 0.07
MMB-by-MMB0.35 ± 0.370.36 ± 0.370.36 ± 0.372.65 ± 1.501.12 ± 0.700.74 ± 0.68
LatentLCD0.56 ± 0.280.60 ± 0.260.58 ± 0.272.30 ± 2.080.51 ± 0.413.94 ± 3.31
MB-by-MB0.27 ± 0.300.39 ± 0.350.31 ± 0.315.95 ± 3.600.67 ± 0.353.05 ± 1.83
CMB0.38 ± 0.370.40 ± 0.380.39 ± 0.372.90 ± 1.871.28 ± 1.246.72 ± 7.20
LoCL (ours)0.67 ± 0.290.68 ± 0.310.67 ± 0.291.40 ± 1.200.03 ± 0.090.06 ± 0.05
Sample size n=5000n=5000
σ-FCI0.55 ± 0.320.61 ± 0.290.57 ± 0.302.55 ± 2.250.54 ± 0.2312.33 ± 6.58
L-MARVEL0.64 ± 0.360.71 ± 0.380.67 ± 0.371.85 ± 2.150.08 ± 0.025.00 ± 0.18
MMB-by-MMB0.49 ± 0.390.46 ± 0.370.47 ± 0.382.15 ± 1.500.84 ± 0.390.45 ± 0.56
LatentLCD0.49 ± 0.320.58 ± 0.300.52 ± 0.313.10 ± 2.430.43 ± 0.373.20 ± 2.74
MB-by-MB0.20 ± 0.190.43 ± 0.330.26 ± 0.238.20 ± 4.691.06 ± 0.503.44 ± 1.81
CMB0.42 ± 0.350.49 ± 0.380.45 ± 0.353.25 ± 2.472.49 ± 1.888.60 ± 7.68
LoCL (ours)0.86 ± 0.240.84 ± 0.250.85 ± 0.250.65 ± 1.010.03 ± 0.090.09 ± 0.13

Table R4. Nonlinear Uniform models.

MethodPrecision ↑Recall ↑F1 ↑Local SHD ↓Runtime (s) ↓CI tests (×10³) ↓
Sample size n=1000n=1000
σ-FCI0.41 ± 0.200.45 ± 0.190.43 ± 0.192.90 ± 1.480.37 ± 0.087.32 ± 1.05
L-MARVEL0.32 ± 0.300.36 ± 0.330.34 ± 0.313.65 ± 2.080.04 ± 0.004.67 ± 0.04
MMB-by-MMB0.40 ± 0.260.38 ± 0.270.39 ± 0.262.47 ± 1.061.24 ± 0.220.92 ± 0.40
LatentLCD0.36 ± 0.170.43 ± 0.180.38 ± 0.173.50 ± 1.660.42 ± 0.262.71 ± 2.40
MB-by-MB0.25 ± 0.230.35 ± 0.260.28 ± 0.235.70 ± 4.460.92 ± 0.957.85 ± 10.34
CMB0.30 ± 0.280.30 ± 0.280.30 ± 0.283.60 ± 2.132.06 ± 1.9017.03 ± 15.21
LoCL (ours)0.46 ± 0.240.46 ± 0.240.46 ± 0.242.15 ± 0.960.03 ± 0.090.04 ± 0.03
Sample size n=3000n=3000
σ-FCI0.55 ± 0.210.61 ± 0.170.58 ± 0.192.25 ± 1.340.45 ± 0.129.01 ± 2.47
L-MARVEL0.67 ± 0.300.74 ± 0.280.69 ± 0.291.85 ± 1.930.06 ± 0.014.85 ± 0.11
MMB-by-MMB0.54 ± 0.300.57 ± 0.320.55 ± 0.301.87 ± 1.190.81 ± 0.630.63 ± 0.64
LatentLCD0.51 ± 0.230.59 ± 0.180.54 ± 0.212.65 ± 1.850.54 ± 0.344.45 ± 2.82
MB-by-MB0.22 ± 0.230.38 ± 0.290.26 ± 0.257.10 ± 3.870.78 ± 0.383.52 ± 2.24
CMB0.37 ± 0.350.43 ± 0.390.39 ± 0.362.90 ± 1.641.29 ± 0.886.74 ± 5.45
LoCL (ours)0.73 ± 0.250.74 ± 0.240.73 ± 0.251.15 ± 1.150.03 ± 0.090.09 ± 0.11
Sample size n=5000n=5000
σ-FCI0.65 ± 0.320.73 ± 0.260.68 ± 0.292.00 ± 2.020.64 ± 0.3614.06 ± 8.66
L-MARVEL0.71 ± 0.320.78 ± 0.310.74 ± 0.311.60 ± 1.930.08 ± 0.015.01 ± 0.13
MMB-by-MMB0.55 ± 0.390.55 ± 0.390.55 ± 0.391.80 ± 1.570.83 ± 0.510.64 ± 0.60
LatentLCD0.60 ± 0.320.66 ± 0.280.62 ± 0.302.25 ± 2.050.58 ± 0.444.00 ± 3.32
MB-by-MB0.19 ± 0.200.36 ± 0.300.24 ± 0.238.25 ± 4.741.22 ± 0.753.54 ± 2.58
CMB0.50 ± 0.320.54 ± 0.340.51 ± 0.332.35 ± 1.593.21 ± 4.7913.51 ± 23.64
LoCL (ours)0.83 ± 0.290.83 ± 0.290.83 ± 0.290.70 ± 1.140.03 ± 0.090.07 ± 0.07

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Main Revisions

We thank the reviewers and the Area Chair for their time and constructive feedback. The table below summarizes the main revisions and their locations in the revised manuscript PDF. The corresponding changes are highlighted in blue for ease of reference. Reviewer IDs and response labels (A1, A2, etc.) identify the corresponding answers in our point-by-point responses.

Main revisionReviewer concernLocation in revised PDF
Added sensitivity experiments on σ-faithfulness violations and discussionsmV61-A1; Wm1c-A1, A4App. E.7, Table 14; App. F
Added experiments with linear non-Gaussian and nonlinear feedback models.mV61-A2; Wm1c-A2, A4Apps. E.4–E.5; Tables 9–12
Added experiments on larger graphs with 500 nodes.Wm1c-A2Sec. 6.1; Fig. 3; App. E.3, Table 8
Added experiments with varying numbers of latent variables.Wm1c-A2App. E.6; Table 13
Added experiments on the acyclic ANDES network.Wm1c-A3Sec. 6.3; Table 2
Added sensitivity experiments on MB estimation errors.Wm1c-A4, A5App. E.8; Table 15
Added MMB-by-MMB as a baseline.G48n-A1Sec. 6; Tables 1, 2, and 5–13
Added visual comparisons of recovered and ground-truth local structures.G48n-A2App. E.9, Figs. 6–7

We hope these revisions and our detailed responses address the reviewers’ concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.