Pith. sign in

REVIEW 5 minor 97 references

Integrating Background Knowledge for Scalable Causal Discovery

T0 review · 0 major / 5 minor · reviewed 2026-07-14 · grok-4.5

Pith's one-line read Background knowledge used during causal discovery can cut CI tests by an order of magnitude while improving effect estimates.

desk verdict Solid engineering paper: careful during-search BK (especially delayed gap handling) for PC/SNAP/local methods, with proofs and clear CI/runtime gains under consistent BK. read the letter →

arxiv 2607.10456 v1 pith:PZIZMA3L submitted 2026-07-11 stat.ML cs.LG

classification stat.MLcs.LG
keywords causaldiscoverybackgroundknowledgePCalgorithmSNAPMarkovblanketMPDAGconditionalindependencetestslocal
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Causal discovery from data alone becomes expensive as the number of variables grows, and most pipelines only apply expert knowledge after the fact to tidy the learned graph. This paper shows that knowledge of which edges exist, which are missing, and which are oriented can be injected into the search itself, pruning candidate conditioning sets and skipping known adjacencies. The authors build this idea into both global algorithms (PC, SNAP) and local algorithms that recover only the structure around a few targets (MB-by-MB, LDECC, LOAD), plus Markov-blanket discovery. Under oracle conditional-independence tests and consistent knowledge they prove the modified procedures remain sound and, with one exception, complete. Empirically, both the number of tests and wall-clock time fall as more knowledge is supplied, while the accuracy of downstream causal-effect estimates stays the same or improves. The practical payoff is that graphs with hundreds of variables become tractable once modest expert constraints are available.

What carries the argument

The three operating principles (skip known adjacencies, restrict separators to possible parents given the current skeleton and BK, and search for separators of known gaps only after the first skeleton pass) together with the b-possible-ancestral pruning rule that lets SNAP keep only the variables still needed for the targets.

What would settle it

Run the same suite of algorithms on graphs with deliberately injected BK errors (as in the paper’s own App. J) and check whether the claimed reductions in CI-test count and intervention distance still appear once the error rate exceeds roughly 10–20 percent.

Watch

Extended reading notes

Core claim

When background knowledge about direct causal relations is used inside the discovery loop—skipping known adjacencies, restricting candidate separators to possible parents, and delaying gap tests until the skeleton is sparse—the resulting algorithms recover the same (or a correctly restricted) MPDAG or local structure while performing substantially fewer conditional-independence tests.

Load-bearing premise

The supplied background knowledge never contradicts the true causal graph; if it does, edges may be kept or mis-oriented and some algorithms become slower rather than faster.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

0 major / 5 minor

Summary. The paper develops a framework for integrating background knowledge (BK) about direct causal relations (adjacencies, orientations, and gaps) into the causal discovery process itself, rather than only as post-processing. It focuses on scalable methods that recover only relevant subgraphs or local structure. The authors instantiate the framework for PC, SNAP(k)/SNAP(∞), MB-by-MB, LDECC, LOAD, and two Markov-blanket algorithms (Grow-Shrink, Total Conditioning). Core design choices include skipping CI tests for known adjacencies, restricting candidate separators to possible parents under BK, and delaying (rather than naively removing) known gaps so that separating sets remain well-defined for v-structure orientation. Under oracle CI tests and consistent BK they prove soundness for all variants and completeness for all except LDECC-BK (Thms. 3.1–3.4, Cor. 3.1–3.2, App. C). Experiments on Erdős–Rényi graphs and bnlearn networks show reductions in CI tests and runtime, with improved or preserved intervention distance as the BK rate rises (Fig. 1, Apps. G–K), plus ablations on BK type, imperfect BK, and post-processing baselines.

Significance. The work addresses a genuine practical bottleneck: constraint-based discovery becomes expensive at scale, yet expert or experimental BK is often available and currently under-used during search. The technical contribution is concrete and reusable—shared subroutines (SkelStep-BK, MissingSepsets-BK, PossPa, PossMB) plus algorithm-specific adaptations that avoid known failure modes of library implementations (App. D). Soundness/completeness proofs with counter-examples for incorrect designs (Ex. C.1–C.4), reproducible code, and multi-metric evaluation (CI counts, time, intervention distance, SHD for global methods) strengthen the claim. The imperfect-BK ablations (App. J) and post-processing comparisons (App. H) make the scope of the claims transparent. If the results hold, the framework is immediately useful for practitioners who already run PC/SNAP/local methods and have partial edge knowledge.

minor comments (5)
  1. In Sec. 3.1 the definition of PossPa_G(X,B) is given inline; a numbered equation would make later references (e.g., the adjusted PossPa* in Eq. (1) for MB-by-MB) cleaner.
  2. Fig. 1 packs five algorithms × three data regimes; the caption and legend are dense. A short table of relative CI-test reductions at BK rate 0.5 would help readers extract the main quantitative message.
  3. App. E.2 (FCI-BK) is preliminary and correctly labeled as such; a one-sentence pointer in the main conclusions that latent-variable completeness remains open would set expectations more clearly for readers who skip the appendix.
  4. A few typos: “SkelStep-BK” vs “SkelStep” consistency in Alg. 4 line 5; “WaiList” in Ex. C.3; “or all CI tests” (missing “f”) in Sec. 5.
  5. Related work could briefly note how the delayed-gap strategy differs from fixedGaps in pcalg/causal-learn/pgmpy (already illustrated in App. D) so that the novelty relative to existing library options is explicit in the main text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: algorithmic soundness under stated assumptions plus empirical comparisons; no fitted inputs re-labeled as predictions.

full rationale

The paper's central claims are (i) soundness/completeness theorems for BK-aware variants of PC, SNAP, MB-by-MB, LOAD, and GrowShrink under oracle CI tests and consistent background knowledge B, and (ii) empirical reductions in CI tests/runtime with non-worsening intervention distance as the BK rate rises. The theorems are proved by adapting standard PC/RFCI/Meek arguments and prior SNAP/LOAD lemmas (App. C); they do not define the target MPDAG or local structure in terms of the algorithm's own output. Empirical intervention distance is measured on held-out samples of size 10000, not on quantities fitted from the discovery data. Self-citations to SNAP and LOAD are prior methods being extended with BK subroutines (SkelStep-BK, MissingSepsets-BK, PossPa), not uniqueness theorems or ansätze that force the new results. Gap-handling is motivated by concrete library failures (App. D) and covered by the soundness proofs rather than by redefining non-adjacency. Imperfect-BK ablations (App. J) are reported as degradation, not as predictions. No step reduces a claimed prediction or first-principles result to its inputs by construction. Score 0 is therefore appropriate.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The work sits on standard causal-discovery assumptions plus the modeling choice that BK is pairwise direct-relation knowledge and consistent. Free parameters are experimental knobs (α, graph ensemble, BK rates), not physical constants fitted into a theory. No new ontological entities are postulated.

free parameters (4)
  • CI significance level α = 0.01
    Fixed at 0.01 for Fisher-Z and G² tests; affects finite-sample skeleton and thus empirical claims.
  • BK sampling rate and type mix = 0–1 rate; uniform adj/orient when adjacent
    Fraction of pairs given BK and random mix of adjacency/orientation/gap; controls the main experimental axis.
  • Erdős–Rényi graph ensemble = n=100, E[deg]=3, maxdeg=10
    n=100, expected degree 3, max degree 10, edge-weight ranges for linear Gaussian; defines the synthetic regime.
  • Imperfect-BK error rates = 10% adj/orient, 5% gaps (main imperfect); 0–30% sweep
    Ablation injects 10%/5% (or 10–30%) errors into adj/orient/gaps; chosen by authors to stress robustness.
assumptions (5)
  • domain assumption Observational distribution p is Markov and faithful to the true causal DAG D.
    Stated in Sec. 2; equates CI tests with d-separations for all constraint-based steps.
  • domain assumption No latent confounders or selection bias in the main contributions.
    Sec. 2; App. E.2 only sketches limited FCI-BK. Completeness claims do not cover PAGs.
  • domain assumption Background knowledge B is consistent with the true DAG D.
    Sec. 2 and all main theorems; imperfect BK is only empirical (App. J).
  • standard math If X and Y are non-adjacent then they are d-separated by Pa(X) or Pa(Y) (Spirtes et al. Lem. 3.3.9).
    Used to justify restricting candidate separating sets to PossPa under BK.
  • standard math Meek orientation rules and MPDAG theory of Perković et al. correctly maximize orientations given BK.
    Alg. 12 and post-orientation steps; completeness of PC-BK/SNAP(∞)-BK rests on this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Integrating Background Knowledge for Scalable Causal Discovery." pith.science (2026). https://pith.science/paper/PZIZMA3L

@misc{pith2026260710456,
  author       = {Pith},
  title        = {Pith review of: Integrating Background Knowledge for Scalable Causal Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PZIZMA3L}},
  note         = {Machine review of arXiv:2607.10456}
}
read the original abstract

Expert background knowledge is often available in practical applications of causal discovery. Such constraints on the true causal graph can help causal discovery in terms of identifiability of causal effects and accuracy of the learned structure, but also in reducing the space of candidate causal graphs. As causal discovery can become computationally expensive for large number of variables, it is crucial to utilize background knowledge effectively during the causal discovery process. However, most current methods only use background knowledge in a postprocessing step after causal discovery to refine the learned graph. In this work, we develop a framework for utilizing background knowledge during the causal discovery process, focusing especially on scalable causal discovery methods that recover only a subset of the whole graph. We implement our framework for multiple algorithms and empirically show that utilizing background knowledge can both reduce computational requirements and increase the quality of the learned structures.

Figures

Figures reproduced from arXiv: 2607.10456 by the authors.

Figure 1
Figure 1. Results over rate of BK between variable pairs. The shadow area denotes standard deviation. Our results in [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Fig. 2a: Example DAG with target T where utilizing background knowledge B, denoted in red, and naively pruning variables that are not b-possible ancestors of T can lead to incorrect results in SNAP(k)-BK. Fig. 2b: The skeleton discovered at order 0. Fig. 2c: The graph after orienting v-structures and orienting edges according to B. Fig. 2d: The graph after pruning the variables that are not b-possible ancestors of T… view at source ↗
Figure 3
Figure 3. Example DAGs where utilizing B, denoted in red, during skeleton search can lead to incorrect results in SNAP. In both cases, the skeleton step up to order 1 never finds V1 ⊥⊥ V3|V2 due to the utilization of background knowledge, keeping V1 and V3 adjacent. On the other hand, it may find V0 ⊥⊥ V1|V2 before V0 ⊥⊥ V1|V3. In that case, at order 1, SNAP would orient a v-structure V1 → V3 ← V0, which contradicts V3 → V0 i… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: The example DAG/MPDAG shown in Fig. 4a with [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]
Figure 5
Figure 5. Figure 5: Results over rate of BK between variable pairs for FCI compared to a preliminary version of FCI-BK as described in App. E.2, on random graphs with an average degree of 3, 40 observed nodes and average 4.4 latent confounders. The shadow area denotes standard deviation. …
Figure 6
Figure 6. Figure 6: SHD for PC-BK and SNAP(∞)-BK over rate of BK between variable pairs. The shadow area denotes standard deviation. G Results on Structural Hamming Distance The various methods we extend recover different types of subgraphs with different sets of nodes or even just local …
Figure 7
Figure 7. Figure 7: Results for showing the improvement of our approach compared to the post-processing baselines described in App. H, quantified as the difference between the baselines and our approach, in the same setting as [PITH_FULL_IMAGE:figures/full_fig_p030_7.png]
Figure 8
Figure 8. Figure 8: Results comparing our approach to post-processing baselines described in App. H under [PITH_FULL_IMAGE:figures/full_fig_p031_8.png]
Figure 9
Figure 9. Figure 9: Results for the PC-BK algorithm over rates of different types background knowledge [PITH_FULL_IMAGE:figures/full_fig_p032_9.png]
Figure 10
Figure 10. Figure 10: Results for the SNAP(∞)-BK algorithm over rates of different types background knowl￾edge between variable pairs. SNAP(∞)-BK, degrades gradually rather than sharply, demonstrating that they are resilient against higher rates of BK error. In our experiments shown in [P…
Figure 11
Figure 11. Figure 11: Results for the MB-by-MB+-BK algorithm over rates of different types background knowledge between variable pairs. The largest linear Gaussian causal graph is ARTH150 with 107 variables 150 edges. Without any background knowledge, the original PC and LDECC algorithms (…
Figure 12
Figure 12. Figure 12: Results for the LDECC+-BK algorithm over rates of different types background knowledge between variable pairs. LOAD-BK algorithm d-separation tests 0.0 0.2 0.4 0.6 0.8 1.0 Rate of background knowledge 1000 1250 1500 1750 2000 2250 2500 Number of CI tests 0.0 0.2 0.4 0…
Figure 13
Figure 13. Figure 13: Results for the LOAD-BK algorithm over rates of different types background knowledge [PITH_FULL_IMAGE:figures/full_fig_p035_13.png]
Figure 14
Figure 14. Figure 14: Results over rate of BK between variable pairs, where BK contains errors at a fixed rate of 10% for adjacencies and orientations, and 5% for gaps. The shadow area denotes standard deviation. 36 [PITH_FULL_IMAGE:figures/full_fig_p036_14.png]
Figure 15
Figure 15. Figure 15: Results over rate of BK error, with BK rate kept fixed at 0.2. The error rate for gaps is always a half of the reported number (e.g., at a rate of 0.1, BK contains errors for 10% of adjacencies and orientations, and for 5% of gaps) to keep the errors balanced between …
Figure 16
Figure 16. Figure 16: Results on semi-synthetic data according to various causal graphs from the [PITH_FULL_IMAGE:figures/full_fig_p038_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

97 extracted references · 7 canonical work pages

  1. [1]

    Cambridge university press, 2009

    Judea Pearl.Causality. Cambridge university press, 2009

  2. [2]

    Review of causal discovery methods based on graphical models.Frontiers in Genetics, 10, 2019

    Clark Glymour, Kun Zhang, and Peter Spirtes. Review of causal discovery methods based on graphical models.Frontiers in Genetics, 10, 2019. ISSN 1664-8021. doi: 10.3389/fgene.2019. 00524

  3. [3]

    Maathuis

    Emilija Perkovi´c, Johannes Textor, Markus Kalisch, and Marloes H. Maathuis. Complete graphical characterization and construction of adjustment sets in markov equivalence classes of ancestral graphs.Journal of Machine Learning Research, 18(220):1–62, 2018. URL http://jmlr.org/papers/v18/16-319.html

  4. [4]

    Inferring causation from time series in earth system sciences.Nature communications, 10(1):2553, 2019

    Jakob Runge, Sebastian Bathiany, Erik Bollt, Gustau Camps-Valls, Dim Coumou, Ethan Deyle, Clark Glymour, Marlene Kretschmer, Miguel D Mahecha, Jordi Mu˜noz-Mar´ı, et al. Inferring causation from time series in earth system sciences.Nature communications, 10(1):2553, 2019

  5. [5]

    Local causal discovery for structural evidence of direct discrimination.Proceedings of the AAAI Conference on Artificial Intelligence, 39(18):19349–19357, Apr

    Jacqueline Maasch, Kyra Gan, Violet Chen, Agni Orfanoudaki, Nil-Jana Akpinar, and Fei Wang. Local causal discovery for structural evidence of direct discrimination.Proceedings of the AAAI Conference on Artificial Intelligence, 39(18):19349–19357, Apr. 2025. doi: 10.1609/aaai. v39i18.34130. URLhttps://ojs.aaai.org/index.php/AAAI/article/view/34130

  6. [6]

    Causal discovery under a confounder blanket

    David S Watson and Ricardo Silva. Causal discovery under a confounder blanket. InUncertainty in Artificial Intelligence, pages 2096–2106. PMLR, 2022

  7. [7]

    Jacqueline R. M. A. Maasch, Weishen Pan, Shantanu Gupta, V olodymyr Kuleshov, Kyra Gan, and Fei Wang. Local discovery by partitioning: Polynomial-time causal discovery around exposure-outcome pairs. InThe 40th Conference on Uncertainty in Artificial Intelligence, 2024

  8. [8]

    Snap: Sequential non-ancestor pruning for targeted causal effect estimation with an unknown graph

    M´aty´as Schubert, Tom Claassen, and Sara Magliacane. Snap: Sequential non-ancestor pruning for targeted causal effect estimation with an unknown graph. In Yingzhen Li, Stephan Mandt, Shipra Agrawal, and Emtiyaz Khan, editors,Proceedings of The 28th International Conference on Artificial Intelligence and Statistics, volume 258 ofProceedings of Machine Lea...

Show all 97 references
  1. [9]

    Discovering and orienting the edges connected to a target variable in a dag via a sequential local learning approach.Computational statistics & data analysis, 77:252–266, 2014

    Changzhang Wang, You Zhou, Qiang Zhao, and Zhi Geng. Discovering and orienting the edges connected to a target variable in a dag via a sequential local learning approach.Computational statistics & data analysis, 77:252–266, 2014

  2. [10]

    Local causal discovery for estimating causal effects

    Shantanu Gupta, David Childers, and Zachary Chase Lipton. Local causal discovery for estimating causal effects. InConference on Causal Learning and Reasoning, pages 408–447. PMLR, 2023

  3. [11]

    Local causal discovery for statistically efficient causal inference

    M´aty´as Schubert, Tom Claassen, and Sara Magliacane. Local causal discovery for statistically efficient causal inference. InThe 29th International Conference on Artificial Intelligence and Statistics, 2026. URLhttps://openreview.net/forum?id=FlWl20PFd7

  4. [12]

    Maathuis

    Emilija Perkovi´c, Markus Kalisch, and Marloes H. Maathuis. Interpreting and using cpdags with background knowledge. In Gal Elidan, Kristian Kersting, and Alexander Ihler, editors, Proceedings of the Thirty-Third Conference on Uncertainty in Artificial Intelligence, UAI 2017, ...

  5. [13]

    On the representation of pairwise causal background knowledge and its applications in causal inference.Journal of Machine Learning Research, 26(229):1–73, 2025

    Zhuangyan Fang, Ruiqi Zhao, Yue Liu, and Yangbo He. On the representation of pairwise causal background knowledge and its applications in causal inference.Journal of Machine Learning Research, 26(229):1–73, 2025. URLhttp://jmlr.org/papers/v26/23-0624.html

  6. [14]

    Do we become wiser with time? on causal equivalence with tiered background knowledge

    Christine W Bang and Vanessa Didelez. Do we become wiser with time? on causal equivalence with tiered background knowledge. InUncertainty in Artificial Intelligence, pages 119–129. PMLR, 2023

  7. [15]

    Ancestral causal inference.Advances in Neural Information Processing Systems, 29, 2016

    Sara Magliacane, Tom Claassen, and Joris M Mooij. Ancestral causal inference.Advances in Neural Information Processing Systems, 29, 2016. 10

  8. [16]

    Sample efficient active learning of causal trees

    Kristjan Greenewald, Dmitriy Katz, Karthikeyan Shanmugam, Sara Magliacane, Murat Ko- caoglu, Enric Boix Adsera, and Guy Bresler. Sample efficient active learning of causal trees. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch ´e-Buc, E. Fox, and R. Garnett, ed- itors,...

  9. [17]

    Large language models for causal discovery: Current landscape and future directions

    Guangya Wan, Yunsheng Lu, Yuqi Wu, Mengxuan Hu, and Sheng Li. Large language models for causal discovery: Current landscape and future directions. In James Kwok, editor,Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25, pages ...

  10. [18]

    On the reliability of large language models for causal discovery

    Tao Feng, Lizhen Qu, Niket Tandon, Zhuang Li, Xiaoxi Kang, and Gholamreza Haffari. On the reliability of large language models for causal discovery. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9565–9590, 2025

  11. [19]

    Prior-knowledge-driven local causal structure learning and its application on causal discovery between type 2 diabetes and bone mineral density.IEEE Access, 8:108798–108810,

    Wei Wang, Gangqiang Hu, Bo Yuan, Shandong Ye, Chao Chen, Yayun Cui, Xi Zhang, and Liting Qian. Prior-knowledge-driven local causal structure learning and its application on causal discovery between type 2 diabetes and bone mineral density.IEEE Access, 8:108798–108810,

  12. [20]

    doi: 10.1109/ACCESS.2020.2994936

  13. [21]

    Challenges and opportunities with causal discovery algorithms: application to alzheimer’s pathophysiology.Scientific reports, 10(1):2975, 2020

    Xinpeng Shen, Sisi Ma, Prashanthi Vemuri, and Gyorgy Simon. Challenges and opportunities with causal discovery algorithms: application to alzheimer’s pathophysiology.Scientific reports, 10(1):2975, 2020

  14. [22]

    Reisch, D ´enes Moln´ar, Stefaan De Henauw, Luis Moreno, Toomas Veidebaum, Michael Tornaritis, Iris Pigeot, and Vanessa Didelez

    Ronja Foraita, Janine Witte, Claudia B¨ornhorst, Wencke Gwozdz, Valeria Pala, Lauren Lissner, Fabio Lauria, Lucia A. Reisch, D ´enes Moln´ar, Stefaan De Henauw, Luis Moreno, Toomas Veidebaum, Michael Tornaritis, Iris Pigeot, and Vanessa Didelez. A longitudinal causal graph ana...

  15. [23]

    Constraint-based causal discovery: Conflict resolution with answer set programming

    Antti Hyttinen, Frederick Eberhardt, and Matti J¨arvisalo. Constraint-based causal discovery: Conflict resolution with answer set programming. In Jin Tian and {Nevin L.} Zhang, editors, Proceedings of the 30th Conference on Uncertainty in Artificial Intelligence (UAI 2014), pa...

  16. [24]

    Constraint-based causal discovery from multiple interventions over overlapping variable sets.Journal of Machine Learning Research, 16(66): 2147–2205, 2015

    Sofia Triantafillou and Ioannis Tsamardinos. Constraint-based causal discovery from multiple interventions over overlapping variable sets.Journal of Machine Learning Research, 16(66): 2147–2205, 2015. URLhttp://jmlr.org/papers/v16/triantafillou15a.html

  17. [25]

    Mooij, Sara Magliacane, and Tom Claassen

    Joris M. Mooij, Sara Magliacane, and Tom Claassen. Joint causal inference from multiple contexts.Journal of Machine Learning Research, 21(99):1–108, 2020. URL http://jmlr. org/papers/v21/17-123.html

  18. [26]

    MIT Press, 2nd edition, 2000

    Peter Spirtes, Clark Glymour, and Richard Scheines.Causation, Prediction, and Search. MIT Press, 2nd edition, 2000

  19. [27]

    Causal inference and causal explanation with background knowledge

    Christopher Meek. Causal inference and causal explanation with background knowledge. In Proceedings of the Eleventh Conference on Uncertainty in Artificial Intelligence, UAI’95, page 403–410, San Francisco, CA, USA, 1995. Morgan Kaufmann Publishers Inc. ISBN 1558603859

  20. [28]

    Maathuis, and Peter B¨uhlmann

    Markus Kalisch, Martin M¨achler, Diego Colombo, Marloes H. Maathuis, and Peter B¨uhlmann. Causal inference using graphical models with the R package pcalg.Journal of Statistical Software, 47(11):1–26, 2012. doi: 10.18637/jss.v047.i11

  21. [29]

    Causal-learn: Causal discovery in python.Journal of Machine Learning Research, 25(60):1–8, 2024

    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.Journal of Machine Learning Research, 25(60):1–8, 2024. 11

  22. [30]

    Tetrad—a toolbox for causal discovery.8th international workshop on climate informatics, 2018

    Joseph D Ramsey, Kun Zhang, Madelyn Glymour, Ruben Sanchez Romero, Biwei Huang, Imme Ebert-Uphoff, Savini Samarasinghe, Elizabeth A Barnes, and Clark Glymour. Tetrad—a toolbox for causal discovery.8th international workshop on climate informatics, 2018

  23. [31]

    pgmpy: A python toolkit for bayesian networks.Journal of Machine Learning Research, 25(265):1–8, 2024

    Ankur Ankan and Johannes Textor. pgmpy: A python toolkit for bayesian networks.Journal of Machine Learning Research, 25(265):1–8, 2024. URL http://jmlr.org/papers/v25/ 23-0487.html

  24. [32]

    Learning high-dimensional directed acyclic graphs with latent and selection variables.The Annals of Statistics, pages 294–321, 2012

    Diego Colombo, Marloes H Maathuis, Markus Kalisch, and Thomas S Richardson. Learning high-dimensional directed acyclic graphs with latent and selection variables.The Annals of Statistics, pages 294–321, 2012

  25. [33]

    Bayesian network induction via local neighborhoods

    Dimitris Margaritis and Sebastian Thrun. Bayesian network induction via local neighborhoods. Advances in neural information processing systems, 12, 1999

  26. [34]

    Using markov blankets for causal structure learning

    Jean-Philippe Pellet and Andr´e Elisseeff. Using markov blankets for causal structure learning. Journal of Machine Learning Research, 9(7), 2008

  27. [35]

    Local Causal Discovery with Background Knowledge .IEEE Transactions on Pattern Analysis & Machine Intelligence, pages 1– 12, February 2026

    Qingyuan Zheng, Yue Liu, and Yangbo He. Local Causal Discovery with Background Knowledge .IEEE Transactions on Pattern Analysis & Machine Intelligence, pages 1– 12, February 2026. ISSN 1939-3539. doi: 10.1109/TPAMI.2026.3667409. URL https: //doi.ieeecomputersociety.org/10.1109...

  28. [36]

    Improving finite sample performance of causal discovery by exploiting temporal structure.arXiv preprint arXiv:2406.19503, 2024

    Christine W Bang, Janine Witte, Ronja Foraita, and Vanessa Didelez. Improving finite sample performance of causal discovery by exploiting temporal structure.arXiv preprint arXiv:2406.19503, 2024

  29. [37]

    Bryan Andrews, Peter Spirtes, and Gregory F. Cooper. On the completeness of causal discovery in the presence of latent confounding with tiered background knowledge. In Silvia Chiappa and Roberto Calandra, editors,Proceedings of the Twenty Third International Conference on Arti...

  30. [38]

    Constraint-based causal discovery with tiered back- ground knowledge and latent variables in single or overlapping datasets.Proceedings of Machine Learning Research, 275:1–31, 2025

    Christine W Bang and Vanessa Didelez. Constraint-based causal discovery with tiered back- ground knowledge and latent variables in single or overlapping datasets.Proceedings of Machine Learning Research, 275:1–31, 2025

  31. [39]

    Score-based causal discovery with temporal background information.arXiv preprint arXiv:2502.06232, 2025

    Tobias Ellegaard Larsen, Claus Thorn Ekstrøm, and Anne Helby Petersen. Score-based causal discovery with temporal background information.arXiv preprint arXiv:2502.06232, 2025

  32. [40]

    Sound and complete causal identification with latent variables given local background knowledge.Artificial Intelligence, 322:103964,

    Tian-Zuo Wang, Tian Qin, and Zhi-Hua Zhou. Sound and complete causal identification with latent variables given local background knowledge.Artificial Intelligence, 322:103964,

  33. [41]

    doi: https://doi.org/10.1016/j.artint.2023.103964

    ISSN 0004-3702. doi: https://doi.org/10.1016/j.artint.2023.103964. URL https: //www.sciencedirect.com/science/article/pii/S0004370223001108

  34. [42]

    Towards complete causal explanation with expert knowledge.arXiv preprint arXiv:2407.07338, 2024

    Aparajithan Venkateswaran and Emilija Perkovi´c. Towards complete causal explanation with expert knowledge.arXiv preprint arXiv:2407.07338, 2024

  35. [43]

    A logical characterization of constraint-based causal discovery

    Tom Claassen and Tom Heskes. A logical characterization of constraint-based causal discovery. InProceedings of the Twenty-Seventh Conference on Uncertainty in Artificial Intelligence, pages 135–144, 2011

  36. [44]

    A hybrid algorithm for learning causal networks using uncertain experts’ knowledge

    Christophe Gonzales, Axel Journe, and Ahmed Mabrouk. A hybrid algorithm for learning causal networks using uncertain experts’ knowledge. In Antonio Salmer´on and Rafael Rum´ı, editors,Proceedings of The 11th International Conference on Probabilistic Graphical Models, volume 18...

  37. [45]

    URLhttps://proceedings.mlr.press/v186/gonzales22a.html

  38. [46]

    Mitigating prior errors in causal structure learning: A resilient approach via bayesian networks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Lyuzhou Chen, Taiyu Ban, Xiangyu Wang, Derui Lyu, and Huanhuan Chen. Mitigating prior errors in causal structure learning: A resilient approach via bayesian networks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 12

  39. [47]

    Causal discovery with language models as imperfect experts

    Stephanie Long, Alexandre Pich´e, Valentina Zantedeschi, Tibor Schuster, and Alexandre Drouin. Causal discovery with language models as imperfect experts. InICML 2023 Workshop on Structured Probabilistic Inference{\&}Generative Modeling, 2023

  40. [48]

    Integrating large language models in causal discovery: A statistical causal approach.Transactions on Machine Learning Research, 2025

    Masayuki Takayama, Tadahisa Okuda, Thong Pham, Tatsuyoshi Ikenoue, Shingo Fukuma, Shohei Shimizu, and Akiyoshi Sannai. Integrating large language models in causal discovery: A statistical causal approach.Transactions on Machine Learning Research, 2025

  41. [49]

    Llm-initialized differentiable causal discovery

    Shiv Kampani, David Hidary, Constantijn van der Poel, Martin Ganahl, and Brenda Miao. Llm-initialized differentiable causal discovery. InCausality and Large Models@ NeurIPS 2024, 2024

  42. [50]

    Large- scale hierarchical causal discovery via weak prior knowledge.IEEE Transactions on Knowledge and Data Engineering, 37(5):2695–2711, 2025

    Xiangyu Wang, Taiyu Ban, Lyuzhou Chen, Derui Lyu, Qinrui Zhu, and Huanhuan Chen. Large- scale hierarchical causal discovery via weak prior knowledge.IEEE Transactions on Knowledge and Data Engineering, 37(5):2695–2711, 2025. doi: 10.1109/TKDE.2025.3537832

  43. [51]

    From guess2graph: When and how can unreliable experts safely boost causal discovery in finite samples?arXiv preprint arXiv:2510.14488, 2025

    Sujai Hiremath, Dominik Janzing, Philipp Faller, Patrick Bl ¨obaum, Elke Kirschbaum, Shiva Prasad Kasiviswanathan, and Kyra Gan. From guess2graph: When and how can unreliable experts safely boost causal discovery in finite samples?arXiv preprint arXiv:2510.14488, 2025

  44. [52]

    Paula Gradu, Tijana Zrnic, Yixin Wang, and Michael I. Jordan. Valid inference after causal discovery.Journal of the American Statistical Association, 120(550):1127–1138,

  45. [53]

    URL https://EconPapers.repec.org/RePEc:taf:jnlasa:v:120:y:2025:i: 550:p:1127-1138

  46. [54]

    Learning bayesian networks with the bnlearn R package.Journal of Statistical Software, 35(3):1–22, 2010

    Marco Scutari. Learning bayesian networks with the bnlearn R package.Journal of Statistical Software, 35(3):1–22, 2010. doi: 10.18637/jss.v035.i03

  47. [55]

    Clarendon Press, 1996

    Steffen L Lauritzen.Graphical models, volume 17. Clarendon Press, 1996

  48. [56]

    Variable elimination, graph reduction and the efficient g-formula.Biometrika, 110(3):739–761, 2023

    F Richard Guo, Emilija Perkovi´c, and Andrea Rotnitzky. Variable elimination, graph reduction and the efficient g-formula.Biometrika, 110(3):739–761, 2023

  49. [57]

    Maathuis, and Vanessa Didelez

    Janine Witte, Leonard Henckel, Marloes H. Maathuis, and Vanessa Didelez. On efficient adjustment in causal graphs.Journal of Machine Learning Research, 21(246):1–45, 2020. URL http://jmlr.org/papers/v21/20-175.html

  50. [58]

    A local method for identifying causal relations under markov equivalence.Artificial Intelligence, 305:103669,

    Zhuangyan Fang, Yue Liu, Zhi Geng, Shengyu Zhu, and Yangbo He. A local method for identifying causal relations under markov equivalence.Artificial Intelligence, 305:103669,

  51. [59]

    doi: https://doi.org/10.1016/j.artint.2022.103669

    ISSN 0004-3702. doi: https://doi.org/10.1016/j.artint.2022.103669. URL https: //www.sciencedirect.com/science/article/pii/S0004370222000091

  52. [60]

    The igraph software package for complex network research

    Gabor Csardi and Tamas Nepusz. The igraph software package for complex network research. InterJournal, Complex Systems:1695, 2006

  53. [61]

    Exploring network structure, dynamics, and function using networkx

    Aric Hagberg, Pieter Swart, and Daniel S Chult. Exploring network structure, dynamics, and function using networkx. Technical report, Los Alamos National Lab.(LANL), Los Alamos, NM (United States), 2008. 13 A Markov blanket discovery with Background Knowledge Markov blanket di...

  54. [62]

    At line 1, all CI tests for pairs of(X, Y)∈ Bare skipped, and thus kept adjacent

  55. [63]

    The first difference impacts all pairs (X, Y)∈ B

    At line 2, the candidate sets to separate X from Y are the P ossP aˆG(X,B) =Ad j ˆG(X)\ ({V∈V|X→V∈ B ∨X̸−V∈ B}). The first difference impacts all pairs (X, Y)∈ B . If X−Y∈ B or X→Y∈ B , then X and Y are adjacent in G∗, thus leaving them adjacent without any CI tests keeps thei...

  56. [64]

    2b, since all variables are marginally dependent

    The skeleton step at order 0 produces the graph shown in Fig. 2b, since all variables are marginally dependent

  57. [65]

    The graph after orienting v-structures (none) and Algorithm 1 of [12] is shown in Fig. 2c

  58. [66]

    In this graph, only V2 is a b-possible ancestor of T , hence V1 is pruned, resulting in the graph shown in Fig. 2d

  59. [67]

    At this point, T and V2 can never be separated, resulting in anincorrectfinal MPDAG shown in Fig. 2d. Ex. C.1 shows why keeping only b-possible ancestors of the targets is not enough at lower orders. Instead, similarly to possibly ancestral sets in [8], we define b-possibly an...

  60. [68]

    Then, because Z→X , it follows that Z∈V ∗

    The first rule involves the structure Z→X−Y . Then, because Z→X , it follows that Z∈V ∗

  61. [69]

    Then, because Z→Y , it follows that Z∈V ∗

    The second rule involves the structure X→Z→Y . Then, because Z→Y , it follows that Z∈V ∗

  62. [70]

    Then, due to V→Y←Z, it follows thatV, Z∈V ∗

    The third rule involves the structures V→Y←Z, V−X−Z and X−Y . Then, due to V→Y←Z, it follows thatV, Z∈V ∗

  63. [71]

    Then, because X−VandZ→Y, it follows thatV, Z∈V ∗

    The fourth rule involves the structuresX−Y, X−V→Z and X−Z→Y . Then, because X−VandZ→Y, it follows thatV, Z∈V ∗. Thus, no edge in G|V∗ could have been oriented by orientations between edges involving variables not inV ∗. This means thatG| V∗ =G ′. We have shown earlier that C| ...

  64. [72]

    At order 0, no variables can be made marginally independent

  65. [73]

    At order 1, the skeleton step may find V0 ⊥ ⊥V1|V2 before V0 ⊥ ⊥V1|V3, resulting in sepset(V0, V1) ={V 2}. 3.B indicates that V2 /∈PossPa(V1) and V2 /∈PossPa(V3), and thus V2 is not required to separate V1 from V3 and V1 ⊥ ⊥V3|V2 is never tested, rendering V1 and V3 still adja...

  66. [74]

    At order 1, the v-structure step orients V1 →V 3 ←V 0, since V3 /∈sepset(V0, V1) ={V 2}

  67. [75]

    The reason why using the PC v-structure orientation rules at order i= 1 in SNAP(k)-BK can fail is because the case of orderi= 1in Lem

    At order 1, SNAP(k)-BKincorrectlyconcludes thatV 3 is a definite non-ancestor ofV 0. The reason why using the PC v-structure orientation rules at order i= 1 in SNAP(k)-BK can fail is because the case of orderi= 1in Lem. B.4. of [8] does not hold anymore. We now show in Lem. C....

  68. [76]

    In the beginning of the algorithm WaitList = [T] and ˆG= ({T, X, V 1, V2, V3, V4},∅)

  69. [77]

    In the first iteration, at line 6, MB-by-MB-BK correctly correctly finds M B(T) = {X, V1, V2, V3} 21

  70. [78]

    (b) When trying to separate nodes fromX, PC-BK also does not findX⊥ ⊥T|{V2} because V2 /∈PossPaLT (X,B)due toX̸−V 2 ∈ B

    Then, at line 14, MB-by-MB proceeds to discover the local structure LT over M B+(T) naively using PC-BK: (a) When trying to separate nodes from T , PC-BK does not find T⊥ ⊥X|{V 2} because V2 /∈PossPaLT (T,B)due toT→V 2 ∈ B. (b) When trying to separate nodes fromX, PC-BK also d...

  71. [79]

    MB-by-MB-BK updates ˆG with edges and v-structures containing T in LT at line 15 includingT→X←V 3

  72. [80]

    However, the edge between X and V1 in LT is discarded as it is not part of a v-structure, due to the extra edge betweenTandX

  73. [81]

    MB-by-MB-BK applies background knowledge T→V 2 at line 16, resulting in the graph shown in Fig. 4b

  74. [82]

    Instead, if PC-BK uses the adjusted definition of possible parents defined in Equation(1), then the MB-by-MB-BK algorithm would proceed as follows:

    As all neighbors of T are oriented, MB-by-MB-BK terminates andincorrectlyreturns X as a child ofT. Instead, if PC-BK uses the adjusted definition of possible parents defined in Equation(1), then the MB-by-MB-BK algorithm would proceed as follows:

  75. [83]

    In the beginning of the algorithm WaiList = [T] and ˆG= ({T, X, V 1, V2, V3, V4},∅)

  76. [84]

    In the first iteration, at line 6, MB-by-MB-BK correctly correctly finds M B(T) = {X, V1, V2, V3}

  77. [85]

    (b) When trying to separate nodes from X at i= 1 of the skeleton search of PC-BK, at this point X and V2 are adjacent, and so V2 ∈PossPa ∗ ˆG(X,B) =Ad j ˆG(X)

    Then, at line 14, MB-by-MB proceeds to discover the local structure LT over M B+(T) using PC-BK: (a) When trying to separate nodes from T , PC-BK does not find T⊥ ⊥X|{V 2} because V2 /∈PossPa∗ ˆG(T,B)due toT→V 2 ∈ B. (b) When trying to separate nodes from X at i= 1 of the skel...

  78. [86]

    MB-by-MB updates its global graph with edges and v-structures containing T at line 15, includingT→V 1 ←X

  79. [87]

    MB-by-MB-BK applies background knowledge T→V 2 at line 16, resulting in the graph shown in Fig. 4c

  80. [88]

    Theorem 3.3.Given oracle CI tests and consistent BK, MB-by-MB-BK is sound and complete in identifying the parents, children and siblings of the target

    As all neighbors of T are oriented, MB-by-MB-BK terminates andcorrectlyreturns X as non-adjacent ofT. Theorem 3.3.Given oracle CI tests and consistent BK, MB-by-MB-BK is sound and complete in identifying the parents, children and siblings of the target. Proof. Alg. 6 utilizes ...

  81. [89]

    At line 1, LDECC-BK findsM B(T) ={Y}

  82. [90]

    At line 2, LDECC-BK findsN e(T) ={Y}

  83. [91]

    At line 3, LDECC-BK identifiesCh(T) ={}

  84. [92]

    The loop at line 7 does not find more independences due to X∗ − ∗Y)∈ B and all other CI tests including targetT

  85. [93]

    Note, that LDECC-BK would also return Y as a sibling of T if it did not skip CI tests for X and Y

    Finally, LDECC-BKincorrectlyreturnsYas a sibling ofT. Note, that LDECC-BK would also return Y as a sibling of T if it did not skip CI tests for X and Y . In that case, it would simply find X̸⊥ ⊥Y|∅ and X̸⊥ ⊥Y|{T} , which do not allow it to orient any edges. C.5 Proof for Thm. ...

  86. [94]

    ˆGT,O contains a directed edge Wi →W j if and only if G contains a directed path Wi → · · · →W j on which all non-endpoint nodes are inF,

  87. [95]

    ˆGT,O contains a bi-directed edge W i↔W j if and only if G contains a path, with at least one non-endpoint node, of the form Wi ← · · · →W j on which all non-endpoints are non-colliders and inF,

  88. [96]

    graphNEL

    ˆGT,O contains an undirected edgeW i −W j if and only ifGcontainsW i −W j. The following lemma, analogous to Lem. 4.4 in [11], shows that the modified forbidden projection in LOAD-BK results in an MPDAG in which the optimal adjustment set are the parents of the outcome. Lemma ...

  89. [97]

    Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.