Pith. sign in

REVIEW 3 major objections 6 minor 55 references

Knowledge-Informed Local Causal Discovery of Optimal Adjustment Sets

T0 review · 3 major / 6 minor · reviewed 2026-07-11 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Injecting required edges into local causal search recovers optimal adjustment sets that pure conditional-independence tests leave unidentified.

desk verdict Clean, useful extension of LOAD that folds background knowledge into the local loop and actually improves adjustment-set recovery under data scarcity; theory is standard composition, main soft spot is the perfect-B+ assumption they already flag. read the letter →

arxiv 2607.04447 v1 pith:4HFT2FQI submitted 2026-07-05 cs.LG stat.ML

Seong Woo Ahn , Alessandro Leite , José Lucas De Melo Costa , Fabrice Popineau , Bich-Liên Doan , Arpad Rimmel This is my paper · ORCID
classification cs.LGstat.ML
keywords localcausaldiscoveryoptimaladjustmentsetsbackgroundknowledgeMPDAGMeekruleseffectestimationMarkovequivalence
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

When only a modest number of samples are available relative to the number of variables, local causal discovery often leaves too many edges undirected to recover a valid optimal adjustment set for a treatment-outcome pair. This paper claims that feeding known directed edges directly into the local search loop, re-imposing them after every update and closing under Meek's four orientation rules, shrinks the remaining equivalence class enough to make those adjustment sets identifiable. The resulting knowledge-constrained local graph can answer causal-effect queries that are unreachable from observational conditional independencies alone. Experiments on random graphs and two biological networks show the largest gains precisely when data are scarce or the local neighborhood is dense, and that knowledge concentrated around the query variables yields the most reliable effect estimates. The practical payoff is a way to turn partial expert pathways into better causal estimates without learning the whole graph.

What carries the argument

b-LOAD's LOCALMPDAG procedure: required edges seed the local graph, are re-imposed after every Markov-blanket update, and are closed under Meek's four rules so that the discovery frontier expands or contracts with every orientation; the resulting local MPDAG is then checked for amenability and used to extract the optimal adjustment set.

What would settle it

On a known biological network, replace a growing fraction of the supplied required edges with false ones and measure when b-LOAD's advantage in recovering the true optimal adjustment set and in intervention distance disappears relative to pure LOAD; the paper's own noise sweep already places that collapse near 50 percent corruption.

Watch

Extended reading notes

Core claim

Under sound background knowledge, the b-LOAD procedure monotonically refines the admissible Markov equivalence class into a knowledge-constrained MPDAG and can enlarge the set of identifiable causal queries, enabling recovery of optimal adjustment sets that are not identifiable from observational conditional-independence information alone.

Load-bearing premise

Every required edge supplied as background knowledge is assumed to be a true edge of the unknown causal graph; the formal guarantees and main experiments rest on that correctness.

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

3 major / 6 minor

Summary. The paper proposes b-LOAD, a knowledge-informed extension of LOAD for local discovery of optimal adjustment sets. Background knowledge (required edges B+) is injected into the local structure-learning loop (LOCALMPDAG): it seeds the graph, is re-imposed after every Markov-blanket update, and is propagated with Meek rules R1–R4, with the wait-list recomputed so that BK-induced orientations can expand the frontier. Under causal sufficiency, Markov+faithfulness, perfect CI tests, and sound B+ ⊆ E, Lemma 1 claims local soundness/completeness of the recovered MPDAG neighborhood, and Proposition 1 claims monotone refinement of the equivalence class, improved amenability/identifiability, and preservation (or possible improvement) of the optimal adjustment set. Empirically, on ER graphs across a statistical-power axis Π and on Sachs/DREAM4, b-LOAD improves F1 of the recovered optimal adjustment set and Intervention Distance relative to PC, b-PC, and LOAD, with largest gains at low Π and with locally targeted BK; a preliminary noise study (Appendix A) shows graceful degradation only for modest corruption.

Significance. If the claims hold under the stated assumptions, the paper fills a genuine gap: local discovery of statistically optimal adjustment sets under background knowledge, with formal MPDAG guarantees and a practical algorithm that integrates BK into the search trajectory rather than post-hoc. The composition of Zheng et al.’s local MPDAG soundness, Meek/Perković MPDAG theory, Perković amenability, and Henckel optimal-adjustment criteria is cleanly stated; the 2×2 experimental design (local/global × with/without BK), the Π power metric, and the local-vs-global sampling prescriptions are useful for practitioners. Code is linked, and the noise analysis, though preliminary, is an honest stress test of the perfect-B+ assumption. The contribution is incremental but well-scoped and actionable for data-scarce causal-effect estimation with curated pathway knowledge.

major comments (3)
  1. Section IV-B and Proposition 1 rest on the working assumption B+ ⊆ E (partial but perfect knowledge). The monotone-identifiability claim and the main synthetic/real results are conditioned on this; Appendix A / Figure 7 only shows residual advantage for μ ≤ 0.3 and collapse of the F1 edge near μ = 0.5. The abstract’s phrase “remains beneficial under moderate structural noise” overstates what is shown. Either strengthen the noise evaluation (forbidden edges B−, weighted/soft constraints, or a clear operating boundary) or narrow the abstract/conclusion claims so they match the perfect-B+ regime that the theory actually covers.
  2. Table II and Section VI: under medium/high BK exposure, b-PC (global) often achieves better Intervention Distance than b-LOAD while b-LOAD wins on F1. The paper correctly notes the F1–Intervention-Distance disconnect but still positions b-LOAD as the preferred method for “more reliable causal-effect estimation.” The prescriptive guidance should be tightened: state explicitly for which (Π, ρ, sampling) cells b-LOAD is preferred on the downstream metric the practitioner consumes, and when a global learner + post-hoc BK is competitive or better.
  3. Lemma 1 and Algorithm 1 assume a perfect CI oracle. Finite-sample CI errors are the regime where the paper claims the largest gains (low Π). The BK-protected re-imposition of B+ is a reasonable safeguard, but there is no analysis of how often data-driven orientations conflict with B+ or of the effect of CI error rates on the recovered MPDAG and O⋆. A short sensitivity experiment (varying α or using a noisy CI oracle) would make the finite-sample claims load-bearing rather than aspirational.
minor comments (6)
  1. Figure 1 caption and abstract: “knowledge-constrained partially directed graph” / “MPDAG” — keep terminology consistent (MPDAG is used later).
  2. Table I: “Efficient Estimation” and “Scalable” columns are not defined; a one-line footnote would help.
  3. Equation (4): Π = n / (|V| · d_exp/2) is clear for ER; note that for non-ER biological graphs the denominator is only an approximation (as done for Sachs/DREAM4).
  4. Section V-C: “Oracle ATE” via OLS on the gold-standard parents is a reasonable first-order proxy; state that it assumes linear mechanisms and may understate error under nonlinear biology.
  5. Typos / style: “b-LOADincorporates”, “b-LOADimproves”, missing spaces after method names in several places; “Erd˝os–Rényi” encoding; “Perkovi´c” accents inconsistent.
  6. Related work: Zheng et al. (2026) and Fang & He (2020) are correctly positioned; a sentence on whether forbidden edges or ancestral path constraints (Borboudakis & Tsamardinos) could be plugged into LOCALMPDAG would round out the discussion.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; claims compose standard Meek/MPDAG and LOAD results under explicit external assumptions, with only expected extension-style citation of the base algorithm.

full rationale

The strongest claim (Proposition 1 and Lemma 1) is a standard composition: valid B+ seeds and re-imposes orientations, four-rule Meek closure (Perković et al. 2017; Meek 1995) yields a refined MPDAG, and amenability/optimal-adjustment criteria (Perković et al. 2018; Henckel et al. 2022) are applied to the local subgraph. The proofs in the appendix reduce directly to these external graphical facts plus the soundness of LOCALMPDAG under perfect CI (invoking Theorem 1 of Zheng et al. 2026, a concurrent non-overlapping work). No quantity is defined in terms of the claimed output; no parameter is fitted to data and then re-presented as a prediction; no uniqueness theorem is imported from the present authors. Citation of LOAD (Schubert et al. 2025a) is the expected base-algorithm reference for an extension paper and is not load-bearing for the new monotone-identifiability argument. Empirical F1 and Intervention Distance are scored against held-out ground-truth DAGs and oracle ATEs, not against method-internal quantities. The perfect-B+ working assumption is stated explicitly and stress-tested (Appendix A), not smuggled. Score 1 only for the minor, non-circular self-reference pattern inherent to naming an extension.

Assumptions & free parameters 3 free parameters · 7 assumptions · 2 invented entities

The central claim rests on standard causal-discovery assumptions plus the paper’s deliberate perfect-knowledge modeling choice. No free parameters are fitted to force the theoretical claim; experimental hyperparameters (ρ grid, Π bins, ER degrees) control evaluation design rather than the derivation. Invented entities are algorithmic constructs, not new physical objects.

free parameters (3)
  • BK fraction ρ
    Evaluation design parameter controlling how much of the true edge set is revealed as B+; not fitted to support the theory, but results are reported across a grid ρ∈{0.1…0.9}.
  • Statistical-power bins for Π
    Low/medium/high Π thresholds (<10, 10–100, ≥100) are chosen by the authors to segment tables; they organize results but are not derived.
  • Corruption rate μ (noise analysis)
    Mean fraction of false required edges in Appendix A; used only for robustness stress tests, not for the main claim.
assumptions (7)
  • domain assumption Causal sufficiency: no latent confounders among observed variables.
    Stated in Lemma 1 and Limitations; inherited by LOAD/PC-style local discovery.
  • domain assumption Markov and faithfulness: conditional independencies in the data match d-separation in the true DAG.
    Section II preliminaries; required for soundness of CI-based skeleton and v-structure steps.
  • ad hoc to paper Background knowledge is sound: B+ ⊆ E of the true DAG (partial but perfect knowledge).
    Section IV-B working assumption; load-bearing for Proposition 1 and main experiments. Forbidden edges left to the skeleton step by design.
  • domain assumption Perfect conditional-independence oracle in the theoretical analysis.
    Lemma 1: “Assume all conditional independencies are correctly identified.” Finite-sample CI error is acknowledged empirically but excluded from the proof.
  • standard math Meek rules R1–R4 yield a maximally oriented MPDAG under background constraints (Meek 1995; Perković et al. 2017).
    Used throughout Algorithm 1 steps (c)–(d) and Proposition 1; standard graphical theory.
  • standard math Amenability and optimal-adjustment criteria of Perković et al. (2018) and Henckel et al. (2022) characterize valid/optimal adjustment sets on MPDAGs.
    Algorithm 1 lines 7–10; theory of what b-LOAD extracts once the local MPDAG is obtained.
  • domain assumption Linear Gaussian SEMs for synthetic data generation and oracle ATE construction on real graphs.
    Section V-B and real-data oracle SEM paragraph; limits external validity to nonlinear/non-Gaussian settings noted in Limitations.
invented entities (2)
  • b-LOAD / LOCALMPDAG procedure with BK-protected updates and BK-aware frontier expansion independent evidence
    purpose: Operationalize knowledge-informed local discovery of optimal adjustment sets by seeding, re-imposing, and propagating B+ inside the local loop.
    Algorithmic construct; independent_evidence is empirical (synthetic and biological benchmarks) rather than a new physical object.
  • Statistical power metric Π = n / (|V| · d_exp/2) independent evidence
    purpose: Collapse heterogeneous graph/sample configurations onto one axis of difficulty for reporting.
    Adapted from Kummerfeld et al. (2024); organizational device for experiments, not a new scientific entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Knowledge-Informed Local Causal Discovery of Optimal Adjustment Sets." pith.science (2026). https://pith.science/paper/4HFT2FQI

@misc{pith2026260704447,
  author       = {Pith},
  title        = {Pith review of: Knowledge-Informed Local Causal Discovery of Optimal Adjustment Sets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4HFT2FQI}},
  note         = {Machine review of arXiv:2607.04447}
}
read the original abstract

Local causal discovery is a scalable alternative to global structure learning. However, it can struggle to identify valid adjustment sets in data-scarce settings because of finite-sample uncertainty, incomplete local neighborhoods, and unresolved Markov equivalence. Although many application domains provide structured background knowledge, its integration into local causal discovery remains limited. We propose b-LOAD, a knowledge-informed extension of the LOAD algorithm for local discovery of optimal adjustment sets. b-LOAD incorporates prior edge constraints directly into the local structure-learning procedure and uses Meek's rules to expand the discovery frontier dynamically, yielding a knowledge-constrained partially directed graph over the relevant local subgraph. This strategy helps prevent structurally relevant nodes introduced by prior knowledge from being excluded by local search. We prove that, under sound background knowledge, the procedure monotonically refines the admissible equivalence class and can enlarge the set of identifiable causal queries, enabling recovery of optimal adjustment sets that are not identifiable from observational conditional-independence information alone. Empirically, b-LOAD improves downstream causal effect estimation relative to purely data-driven and standard knowledge-augmented baselines, particularly in data-scarce and structurally complex regimes. Results on real-world biological networks show that locally targeted prior knowledge provides the largest gains and remains beneficial under moderate structural noise. These findings position b-LOAD as a scalable approach for converting fragmented domain knowledge into more reliable causal-effect estimation.

Figures

Figures reproduced from arXiv: 2607.04447 by the authors.

Figure 1
Figure 1. Overview of b-LOAD. (A) Without background knowledge, LOAD returns an unoriented local CPDAG, leaving the causal query potentially non-identifiable. (B) Background constraints B trigger b-LOAD propagation via Meek rules, refining the equivalence class into an MPDAG. (C) These orientations consistently improve optimal-adjustment-set recovery (F1) across the low-Π, medium-Π, and high-Π graph complexity regimes. discov… view at source ↗
Figure 2
Figure 2. Meek’s four orientation rules (Meek, 1995). Blue lines are undirected edges; blue arrows indicate the newly oriented edge implied by each rule. Nodes not adjacent in the precondition pattern are assumed non-adjacent. ancestral graphs (MAGs), and partial ancestral graphs (PAGs). Among these valid adjustment sets, the optimal adjustment set O⋆ minimizes the asymptotic variance of covariate-adjusted estimators (Henckel… view at source ↗
Figure 3
Figure 3. F1 vs. statistical power Π (Equation (4)) for a 2×2 grid of BK fractions ρ ∈ {20%, 40%, 60%, 80%} and method variants (higher is better). b-LOAD (global, orange dashed) and b-LOAD (local, green dashed) overlap at the top of every panel. Shaded regions: 95% CI over 1000 seeds. B. Synthetic Experiments Data generation. We generate random DAGs using the Erdos–Rényi model, providing fine-grained control over ˝ three axe… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Intervention Distance (lower is better) on synthetic ER2 graphs across three orthogonal views: data efficiency ( [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Sachs: F1 (left) and Intervention Distance (right) under [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: DREAM4 In Silico, averaged over five subnetworks: F1 (left) and Intervention Distance (right) under global (top) and local (bottom) BK sampling. b-LOAD lowest Intervention Distance throughout; b-PC competitive on F1 only under local sampling at high ρ. highly competiti…
Figure 7
Figure 7. Figure 7: Noise extremes (µ=0.1 top, µ=0.5 bottom) on the synthetic sweep. F1 (left) and Intervention Distance (right) vs. Π; line shade encodes ρ. Blue: b-LOAD with noised B +; orange: LOAD. at higher edge densities (ER4, ER5, and ER6) were excluded due to prohibitive time and …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 2 linked inside Pith

  1. [1]

    Spirtes and C

    P. Spirtes and C. Glymour and R. Scheines , title =

  2. [2]

    Chickering , title =

    D. Chickering , title =. Journal of Machine Learning Research , volume =

  3. [3]

    2009 , edition =

    Pearl, Judea , title =. 2009 , edition =

    2009

  4. [4]

    1995 , booktitle =

    Meek, Christopher , title =. 1995 , booktitle =

    1995

  5. [5]

    Maathuis , title =

    Emilija Perkovi\'c and Johannes Textor and Markus Kalisch and Marloes H. Maathuis , title =. Journal of Machine Learning Research , year =

  6. [6]

    Graphical Criteria for Efficient Total Effect Estimation Via Adjustment in Causal Linear Models , journal =

    Henckel, Leonard and Perkovi. Graphical Criteria for Efficient Total Effect Estimation Via Adjustment in Causal Linear Models , journal =. 2022 , month =

    2022

  7. [7]

    Frontiers in Genetics , volume=

    Glymour, Clark and Zhang, Kun and Spirtes, Peter , title=. Frontiers in Genetics , volume=

  8. [8]

    2003 , journal =

    Chickering, David Maxwell , title =. 2003 , journal =

    2003

Show all 55 references
  1. [9]

    and Camgoz, Necati Cihan and Bowden, Richard , title =

    Vowels, Matthew J. and Camgoz, Necati Cihan and Bowden, Richard , title =. 2022 , volume =

    2022

  2. [10]

    Maathuis and Markus Kalisch and Peter B

    Marloes H. Maathuis and Markus Kalisch and Peter B. The Annals of Statistics , number =

  3. [11]

    2015 , booktitle =

    Gao, Tian and Ji, Qiang , title =. 2015 , booktitle =

    2015

  4. [12]

    arXiv:1707.02171 , year=

    Interpreting and using CPDAGs with background knowledge , author=. arXiv:1707.02171 , year=

    arXiv

  5. [13]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Local Causal Discovery with Background Knowledge , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  6. [14]

    Dong, Shuyu and Sebag, Michele and Uemura, Kento and Fujii, Akito and Chang, Shuang and Koyanagi, Yusuke and Maruhashi, Koji , booktitle=

  7. [15]

    2nd Conference on Causal Learning and Reasoning , pages =

    Local Causal Discovery for Estimating Causal Effects , author =. 2nd Conference on Causal Learning and Reasoning , pages =. 2023 , editor =

    2023

  8. [16]

    2020 , editor =

    Fang, Zhuangyan and He, Yangbo , booktitle =. 2020 , editor =

    2020

  9. [17]

    2024 , journal =

    Towards Complete Causal Explanation with Expert Knowledge , author =. 2024 , journal =

    2024

  10. [18]

    arXiv:2510.14582 , year=

    Local Causal Discovery for Statistically Efficient Causal Inference , author=. arXiv:2510.14582 , year=

  11. [19]

    Computational statistics & data analysis , volume=

    Discovering and orienting the edges connected to a target variable in a DAG via a sequential local learning approach , author=. Computational statistics & data analysis , volume=

  12. [20]

    International journal of data science and analytics , volume=

    Power analysis for causal discovery , author=. International journal of data science and analytics , volume=

  13. [21]

    Neural computation , volume=

    Structural intervention distance for evaluating causal graphs , author=. Neural computation , volume=

  14. [22]

    2017 , publisher=

    Elements of causal inference: foundations and learning algorithms , author=. 2017 , publisher=

    2017

  15. [23]

    Statistical Science , volume=

    [Bayesian analysis in expert systems]: Comment: Graphical models, causality and intervention , author=. Statistical Science , volume=. 1993 , publisher=

    1993

  16. [24]

    Artificial Intelligence , volume =

    Zhang, Jiji , title =. Artificial Intelligence , volume =

  17. [25]

    and Aliferis, Constantin F

    Tsamardinos, Ioannis and Brown, Laura E. and Aliferis, Constantin F. , title =. Machine Learning , volume =

  18. [26]

    , title =

    Robinson, Robert W. , title =. Combinatorial Mathematics V: 5th Australian Conference , editor =

  19. [27]

    Pellet, Jean-Philippe and Elisseeff, Andr. Using. Journal of Machine Learning Research , volume =

  20. [28]

    and Madigan, David and Perlman, Michael D

    Andersson, Steen A. and Madigan, David and Perlman, Michael D. , title =. The Annals of Statistics , volume =

  21. [29]

    Shimizu, Shohei and Hoyer, Patrik O. and Hyv. A. Journal of Machine Learning Research , volume =

  22. [30]

    and Nolan, Garry P

    Sachs, Karen and Perez, Omar and Pe'er, Dana and Lauffenburger, Douglas A. and Nolan, Garry P. , title =. Science , volume =

  23. [31]

    Nucleic Acids Research , volume =

    Milacic, Marija and Beavers, Deidre and Conley, Patrick and Gong, Chuqiao and Gillespie, Marc and Griss, Johannes and Haw, Robin and Jassal, Bijay and Matthews, Lisa and May, Bruce and Petryszak, Robert and Ragueneau, Eliot and Rothfels, Karen and Sevilla, Cristoffer and Shamo...

  24. [32]

    arXiv:2502.07857 , year =

    Schubert, M. arXiv:2502.07857 , year =

    arXiv

  25. [33]

    and Schaffter, Thomas and Mattiussi, Claudio and Floreano, Dario and Stolovitzky, Gustavo , title =

    Marbach, Daniel and Prill, Robert J. and Schaffter, Thomas and Mattiussi, Claudio and Floreano, Dario and Stolovitzky, Gustavo , title =. Proceedings of the National Academy of Sciences , volume =

  26. [34]

    2018 , publisher=

    The book of why: The new science of cause and effect , author=. 2018 , publisher=

    2018

  27. [35]

    and Statnikov, Alexander and Tsamardinos, Ioannis and Mani, Subramani and Koutsoukos, Xenofon D

    Aliferis, Constantin F. and Statnikov, Alexander and Tsamardinos, Ioannis and Mani, Subramani and Koutsoukos, Xenofon D. , title =. Journal of Machine Learning Research , volume =

  28. [36]

    Journal of Machine Learning Research , volume =

    Hauser, Alain and B\"uhlmann, Peter , title =. Journal of Machine Learning Research , volume =

  29. [37]

    and Meinshausen, Nicolai , title =

    Heinze-Deml, Christina and Maathuis, Marloes H. and Meinshausen, Nicolai , title =. Annual Review of Statistics and Its Application , volume =

  30. [38]

    and Didelez, Vanessa , title =

    Witte, Janine and Henckel, Leonard and Maathuis, Marloes H. and Didelez, Vanessa , title =. Journal of Machine Learning Research , volume =

  31. [39]

    Journal of Machine Learning Research , volume =

    Rotnitzky, Andrea and Smucler, Ezequiel , title =. Journal of Machine Learning Research , volume =

  32. [40]

    Artificial Intelligence , volume =

    van der Zander, Benito and Li\'skiewicz, Maciej and Textor, Johannes , title =. Artificial Intelligence , volume =

  33. [41]

    , title =

    Heckerman, David and Geiger, Dan and Chickering, David M. , title =. Machine Learning , volume =

  34. [42]

    29th International Conference on Machine Learning , year =

    Borboudakis, Giorgos and Tsamardinos, Ioannis , title =. 29th International Conference on Machine Learning , year =

  35. [43]

    36th Conference on Uncertainty in Artificial Intelligence , volume =

    Fang, Zhuangyan and He, Yangbo , title =. 36th Conference on Uncertainty in Artificial Intelligence , volume =

  36. [44]

    37th Conference on Uncertainty in Artificial Intelligence , year =

    Wien\"obst, Marcel and Bannach, Max and Li\'skiewicz, Maciej , title =. 37th Conference on Uncertainty in Artificial Intelligence , year =

  37. [45]

    Biometrika , volume =

    Pearl, Judea , title =. Biometrika , volume =

  38. [46]

    18th National Conference on Artificial Intelligence , year =

    Tian, Jin and Pearl, Judea , title =. 18th National Conference on Artificial Intelligence , year =

  39. [47]

    , title =

    Shpitser, Ilya and VanderWeele, Tyler and Robins, James M. , title =. 26th Conference on Uncertainty in Artificial Intelligence , year =

  40. [48]

    Advances in Neural Information Processing Systems , year =

    Margaritis, Dimitris and Thrun, Sebastian , title =. Advances in Neural Information Processing Systems , year =

  41. [49]

    and Statnikov, Alexander , title =

    Tsamardinos, Ioannis and Aliferis, Constantin F. and Statnikov, Alexander , title =. 16th International FLAIRS Conference , year =

  42. [50]

    and Jette, Morris A

    Yoo, Andy B. and Jette, Morris A. and Grondona, Mark , title =. Lecture Notes in Computer Science , year =

  43. [51]

    The Computational Geometry Algorithms Library , author =

  44. [52]

    Menelaos Karavelas , subtitle =

  45. [53]

    The Computational Geometry Algorithms Library , subtitle =

    Menelaos Karavelas , editor =. The Computational Geometry Algorithms Library , subtitle =

  46. [54]

    The Parmap library , author =

  47. [55]

    Christopher Anderson and Sophia Drossopoulou , title =

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.