pith. sign in

arxiv: 2606.03251 · v1 · pith:H3USRGTKnew · submitted 2026-06-02 · 💻 cs.AI · cs.CV· cs.LG· eess.IV· stat.ML

Do Real-World Datasets Contain Natural Experiments? An Empirical Study Using Causal Feature Selection

Gautam Gare , John Galeotti , Michael Mozer , Deva Ramanan , Nan Rosemary Ke This is my paper

Pith reviewed 2026-06-28 09:55 UTC · model grok-4.3

classification 💻 cs.AI cs.CVcs.LGeess.IVstat.ML
keywords natural experimentscausal discoverycausal feature selectionreal-world datasetsinterventional modelingcausal inferencemodel performance
0
0 comments X

The pith

Real-world datasets contain natural experiments that causal feature selection can exploit to raise model performance.

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

The paper asks whether events that affect some individuals but not others appear as implicit interventions inside ordinary collected datasets. It recovers a causal graph from each dataset, selects features according to the recovered causal links, and checks whether performance rises when the same data is modeled as an intervention rather than a pure observation. Synthetic graphs that do and do not embed such implicit interventions first confirm that the performance difference tracks the presence of natural experiments. The same test is then run on a large collection of real-world datasets. The consistent lift supports the claim that natural experiments exist in practice and can be used to improve models.

Core claim

Recovering the causal graph and selecting features along causal links produces a measurable performance gain when the data is treated as interventional rather than observational; the authors interpret this gain as evidence that natural experiments are present in the real-world collections they examined.

What carries the argument

Causal discovery that recovers a graph, followed by causal feature selection that separates observational from interventional regimes.

If this is right

  • Synthetic graphs containing natural experiments produce the expected performance lift while graphs without them do not.
  • Multiple real-world datasets exhibit the same lift when switched to interventional modeling.
  • Causal feature selection therefore supplies a practical way to improve prediction without collecting new interventional data.
  • The method supplies an operational test for the existence of natural experiments inside any given collection.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Standard ML benchmarks may already embed usable natural experiments that current training pipelines ignore.
  • Dataset audits could routinely include a causal-graph scan to flag hidden interventions before model training begins.
  • Domains with frequent policy changes or localized shocks are the most promising places to search for additional natural experiments.

Load-bearing premise

Any performance lift from interventional modeling is produced by the presence of natural experiments rather than by incidental changes in selected features or model capacity.

What would settle it

A real-world dataset in which the same causal feature selection yields no performance difference, or in which the recovered graph contains no implicit interventions, would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.03251 by Deva Ramanan, Gautam Gare, John Galeotti, Michael Mozer, Nan Rosemary Ke.

Figure 1
Figure 1. Figure 1: (Left) Conventional ML approaches do not account for naturally occurring experiments in data, which can result in biased predictions. Causal approaches which capture naturally-occuring interventions can lead to more accurate predictions; please see Appendix G for a simple illustrative demonstration. (Right) The causal graph used for synthetic Sachs [6] experiments, as in the BnLearn library [7]. In our exp… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of neural-causal vs. classical causal and non-causal feature-selection methods [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The DAG discovered by causal-sk for the Diabetes dataset. Notably, the Insulin feature is excluded from the Markov blanket of the class node. Although this may seem counterintuitive, an exhaustive evaluation of all 255 feature subsets confirms that the optimal feature set also excludes Insulin ( [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The marginalized Mixed Acyclic Graph (MAG) of the Sachs DAG graph after marginalizing [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The DAG discovered by the causal-sk for the Buddy dataset. [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The DAG discovered by the causal-sk for the Gesture-phase dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The figure presents the results of our analysis conducted on a synthetic dataset generated [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
read the original abstract

In nature, events that affect some individuals or groups but not others constitute an implicit intervention and are known as natural experiments. For example, the COVID-19 pandemic was an intervention by the coronavirus on the sub-population infected with COVID. We ask, do natural experiments occur in existing real-world datasets? If yes, how should we treat them? To detect natural experiments in data, we use causal discovery to recover the underlying causal graph and perform feature selection based on causal links. If downstream performance improves by treating the data as interventional rather than observational, we argue that this suggests the dataset contains natural experiments. We first validate this hypothesis by simulating datasets with and without natural experiments using synthetic graphs. We then perform a systematic empirical evaluation on a large suite of real-world datasets. Our results indicate that real-world datasets do contain natural experiments and we can take advantage of those natural experiments to improve model performance using causal inference. Our work represents the initial foray into this area, offering a preliminary exploration within a limited scope.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims that real-world datasets contain natural experiments (implicit interventions affecting subpopulations), which can be detected by applying causal discovery to recover the underlying graph and then performing feature selection on causal links. Synthetic experiments with controlled graphs validate that performance improves when data is treated as interventional rather than observational precisely when natural experiments are present. A systematic evaluation on a large suite of real-world datasets then shows similar performance gains, leading to the conclusion that such datasets do contain natural experiments and that causal inference can be used to exploit them.

Significance. If the central empirical claim holds after isolating the contribution of natural experiments, the work offers a first systematic probe into whether standard observational datasets embed unrecognized interventions and whether causal feature selection can capitalize on them. The synthetic validation protocol is a clear strength, as it directly controls intervention presence and graph structure to test the detection method.

major comments (2)
  1. [Abstract] Abstract (final paragraph) and the description of the real-world evaluation: the reported performance lift is obtained by comparing a pipeline that recovers a causal graph and selects features on causal links against a baseline that uses standard (non-causal) feature selection. Because the two pipelines differ in both the selected feature set and the induced model structure, the delta cannot be unambiguously attributed to the presence of natural experiments rather than to differences in feature selection quality.
  2. [Synthetic experiments] Synthetic experiments section: while the controlled graphs successfully isolate the effect of natural experiments, the real-world suite lacks a corresponding ablation that holds the feature set fixed and varies only whether the modeling treats the data as interventional or observational. Without this control, the cross-dataset claim that performance gains demonstrate natural experiments remains under-supported.
minor comments (1)
  1. The abstract and method description would benefit from an explicit statement of the precise statistical test used to declare a performance improvement significant across the real-world suite.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed review and insightful comments on our manuscript. We agree that strengthening the attribution of performance improvements to natural experiments is important and will revise the paper accordingly to include additional ablations. Our point-by-point responses follow.

read point-by-point responses

  1. Referee: [Abstract] Abstract (final paragraph) and the description of the real-world evaluation: the reported performance lift is obtained by comparing a pipeline that recovers a causal graph and selects features on causal links against a baseline that uses standard (non-causal) feature selection. Because the two pipelines differ in both the selected feature set and the induced model structure, the delta cannot be unambiguously attributed to the presence of natural experiments rather than to differences in feature selection quality.

    Authors: We thank the referee for highlighting this potential confound. The synthetic experiments are designed to isolate the contribution of natural experiments by comparing performance with and without them under the causal pipeline. When natural experiments are absent, no performance gain is observed despite using the same causal feature selection method. This suggests that the gains are attributable to the natural experiments rather than the feature selection procedure alone. Nevertheless, to address the concern directly for real-world data, we will add an ablation that holds the feature set fixed (using the causal selection) and varies only the modeling assumption (interventional vs. observational). revision: yes

  2. Referee: [Synthetic experiments] Synthetic experiments section: while the controlled graphs successfully isolate the effect of natural experiments, the real-world suite lacks a corresponding ablation that holds the feature set fixed and varies only whether the modeling treats the data as interventional or observational. Without this control, the cross-dataset claim that performance gains demonstrate natural experiments remains under-supported.

    Authors: We concur that an ablation holding the feature set constant would provide clearer evidence. The current synthetic validation already performs this isolation by varying the presence of natural experiments. For the real-world evaluation, we will incorporate the suggested ablation in the revised version, applying causal feature selection and then comparing interventional and observational modeling on the fixed feature set across the datasets. This will allow us to more confidently attribute any gains to the natural experiments. revision: yes

Circularity Check

0 steps flagged

Empirical performance comparison with no derivation chain or self-referential reduction

full rationale

The paper contains no equations, derivations, or first-principles results. Its central claim rests on an empirical protocol: run causal discovery + causal-link feature selection (interventional pipeline) versus standard feature selection (observational baseline), then interpret any accuracy lift on real-world data as evidence of natural experiments. This protocol is validated on controlled synthetics where graph and intervention presence are known, then applied to real datasets. No step reduces a claimed output to its own fitted inputs or to a self-citation chain by construction; the interpretive step (lift implies natural experiments) is a substantive modeling choice, not a definitional tautology. Minor self-citations to causal-discovery literature are present but not load-bearing for the empirical result. The paper is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the method implicitly relies on the correctness of off-the-shelf causal discovery algorithms whose assumptions are not audited here.

pith-pipeline@v0.9.1-grok · 5728 in / 1151 out tokens · 16429 ms · 2026-06-28T09:55:58.469789+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

71 extracted references · 28 canonical work pages · 5 internal anchors

  1. [1]

    Natural experiments: A Nobel Prize awarded research design for strengthening causal inference on global health challenges.Journal of Global Health, 12, 2022

    Famke JM Mölenberg and Francisca Vargas Lopes. Natural experiments: A Nobel Prize awarded research design for strengthening causal inference on global health challenges.Journal of Global Health, 12, 2022. ISSN 20472986. doi: 10.7189/JOGH.12.03005. URL /pmc/articles/ PMC9031504/https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9031504/

    work page doi:10.7189/jogh.12.03005 2022

  2. [2]

    Causal inference in statistics: An overview.Statistics surveys, 3:96–146, 2009

    Judea Pearl et al. Causal inference in statistics: An overview.Statistics surveys, 3:96–146, 2009

    2009

  3. [3]

    Gianicolo, Martin Eichler, Oliver Muensterer, Konstantin Strauch, and Maria Blettner

    Emilio A.L. Gianicolo, Martin Eichler, Oliver Muensterer, Konstantin Strauch, and Maria Blettner. Methods for Evaluating Causality in Observational Studies: Part 27 of a Series on Evaluation of Scientific Publications.Deutsches Ärzteblatt International, 117(7):101, 2 2020. ISSN 18660452. doi: 10.3238/ARZTEBL.2020.0101. URL https://www.ncbi.nlm.nih. gov/pm...

    work page doi:10.3238/arztebl.2020.0101 2020

  4. [4]

    Nobel-winning ‘natural experiments’ approach made economics more robust

    Philip Ball. Nobel-winning ‘natural experiments’ approach made economics more robust. Nature, 10 2021. ISSN 0028-0836. doi: 10.1038/D41586-021-02799-7

    work page doi:10.1038/d41586-021-02799-7 2021

  5. [5]

    Understanding Development and Poverty Alleviation. 2019. URLhttps://www.nobelprize. org/uploads/2019/10/advanced-economicsciencesprize2019.pdf

    2019

  6. [6]

    Causal protein-signaling networks derived from multiparameter single-cell data.Science, 308(5721): 523–529, 2005

    Karen Sachs, Omar Perez, Dana Pe’er, Douglas A Lauffenburger, and Garry P Nolan. Causal protein-signaling networks derived from multiparameter single-cell data.Science, 308(5721): 523–529, 2005

    2005

  7. [7]

    Learning Bayesian Networks with the <b>bnlearn</b> <i>R</i> Package

    Marco Scutari. Learning Bayesian Networks with the <b>bnlearn</b> <i>R</i> Package. Journal of Statistical Software, 35(3):1–22, 7 2010. ISSN 1548-7660. doi: 10.18637/jss.v035. i03. URLhttp://www.jstatsoft.org/v35/i03/

    work page doi:10.18637/jss.v035 2010

  8. [8]

    Differentiable causal discovery from interventional data, 2020 a

    Philippe Brouillard, Sébastien Lachapelle, Alexandre Lacoste, Simon Lacoste-Julien, and Alexandre Drouin. Differentiable Causal Discovery from Interventional Data.Advances in Neural Information Processing Systems, 2020-December, 7 2020. ISSN 10495258. doi: 10.48550/arxiv.2007.01754. URLhttps://arxiv.org/abs/2007.01754v2

    work page doi:10.48550/arxiv.2007.01754 2020

  9. [9]

    Causal diagrams for empirical research.Biometrika, 82(4):669–688, 1995

    Judea Pearl. Causal diagrams for empirical research.Biometrika, 82(4):669–688, 1995

    1995

  10. [10]

    The MIT Press, 2017

    Jonas Peters, Dominik Janzing, and Bernhard Schölkopf.Elements of causal inference: founda- tions and learning algorithms. The MIT Press, 2017

    2017

  11. [11]

    Morgan Kaufmann, San Francisco, Calif., 2009

    Judea Pearl.Probabilistic reasoning in intelligent systems : networks of plausible inference. Morgan Kaufmann, San Francisco, Calif., 2009. ISBN 9781558604797 1558604790. URL https://www.amazon.de/ Probabilistic-Reasoning-Intelligent-Systems-Representation/dp/ 1558604790. Example for Explaning away

    2009

  12. [12]

    Toward optimal feature selection

    Daphne Koller and Mehran Sahami. Toward optimal feature selection. InProc. 13th Interna- tional Conference on Machine Learning, pages 284–292. Morgan Kaufmann, 1996

    1996

  13. [13]

    Aliferis

    Ioannis Tsamardinos and Constantin F. Aliferis. Towards Principled Feature Selection: Rele- vancy, Filters and Wrappers, 1 2003. ISSN 2640-3498. URL https://proceedings.mlr. press/r4/tsamardinos03a.html

    2003

  14. [14]

    Cambridge University Press, USA, 2nd edition,

    Judea Pearl.CAUSALITY, 2nd Edition, 2009. Cambridge University Press, USA, 2nd edition,

    2009

  15. [15]

    URLhttp://bayes.cs.ucla.edu/BOOK-2K/

    ISBN 052189560X. URLhttp://bayes.cs.ucla.edu/BOOK-2K/

  16. [16]

    Chickering

    David Heckerman, Dan Geiger, and David M. Chickering. Learning Bayesian Networks: The Combination of Knowledge and Statistical Data.Machine Learning 1995 20:3, 20(3):197–243,

    1995

  17. [17]

    Chickering

    ISSN 1573-0565. doi: 10.1023/A:1022623210503. URL https://link.springer. com/article/10.1023/A:1022623210503. 11

    work page doi:10.1023/a:1022623210503

  18. [18]

    Spirtes, Clark N

    Peter. Spirtes, Clark N. Glymour, and Richard. Scheines. Causation, prediction, and search. page 543, 2000

    2000

  19. [19]

    Causal discovery with general non-linear relationships using non-linear ica.arXiv preprint arXiv:1904.09096, 2019

    Ricardo Pio Monti, Kun Zhang, and Aapo Hyvarinen. Causal discovery with general non-linear relationships using non-linear ica.arXiv preprint arXiv:1904.09096, 2019

    Pith/arXiv arXiv 1904

  20. [20]

    MIT press, 2000

    Peter Spirtes, Clark N Glymour, Richard Scheines, David Heckerman, Christopher Meek, Gregory Cooper, and Thomas Richardson.Causation, prediction, and search. MIT press, 2000

    2000

  21. [21]

    A kernel-based causal learning algorithm

    Xiaohai Sun, Dominik Janzing, Bernhard Schölkopf, and Kenji Fukumizu. A kernel-based causal learning algorithm. InProceedings of the 24th international conference on Machine learning, pages 855–862. ACM, 2007

    2007

  22. [22]

    Kernel-based conditional independence test and application in causal discovery.arXiv preprint arXiv:1202.3775, 2012

    Kun Zhang, Jonas Peters, Dominik Janzing, and Bernhard Schölkopf. Kernel-based conditional independence test and application in causal discovery.arXiv preprint arXiv:1202.3775, 2012

    Pith/arXiv arXiv 2012

  23. [23]

    Causal Discovery with Reinforcement Learning

    Shengyu Zhu, Ignavier Ng, and Zhitang Chen. Causal Discovery with Reinforcement Learning. 6 2019. doi: 10.48550/arxiv.1906.04477. URLhttp://arxiv.org/abs/1906.04477

    work page doi:10.48550/arxiv.1906.04477 2019

  24. [24]

    Optimal structure identification with greedy search.Journal of machine learning research, 3(Nov):507–554, 2002

    David Maxwell Chickering. Optimal structure identification with greedy search.Journal of machine learning research, 3(Nov):507–554, 2002

    2002

  25. [25]

    Cooper and Changwon Yoo

    Gregory F. Cooper and Changwon Yoo. Causal Discovery from a Mixture of Experimental and Observational Data. InProceedings of the Fifteenth Conference on Uncertainty in Artificial Intelligence, UAI’99, pages 116–125, San Francisco, CA, USA, 1999

    1999

  26. [26]

    Causal generative neural networks.arXiv preprint arXiv:1711.08936, 2017

    Olivier Goudet, Diviyan Kalainathan, Philippe Caillou, Isabelle Guyon, David Lopez-Paz, and Michèle Sebag. Causal generative neural networks.arXiv preprint arXiv:1711.08936, 2017

    Pith/arXiv arXiv 2017

  27. [27]

    Characterization and greedy learning of interventional markov equivalence classes of directed acyclic graphs.The Journal of Machine Learning Research, 13(1):2409–2464, 2012

    Alain Hauser and Peter Bühlmann. Characterization and greedy learning of interventional markov equivalence classes of directed acyclic graphs.The Journal of Machine Learning Research, 13(1):2409–2464, 2012

    2012

  28. [28]

    Learning bayesian networks: The combination of knowledge and statistical data.Machine learning, 20(3):197–243, 1995

    David Heckerman, Dan Geiger, and David M Chickering. Learning bayesian networks: The combination of knowledge and statistical data.Machine learning, 20(3):197–243, 1995

    1995

  29. [29]

    Brown, and Constantin F

    Ioannis Tsamardinos, Laura E. Brown, and Constantin F. Aliferis. The max-min hill-climbing Bayesian network structure learning algorithm.Machine Learning 2006 65:1, 65(1):31–78, 3

    2006

  30. [30]

    doi: 10.1007/S10994-006-6889-7

    ISSN 1573-0565. doi: 10.1007/S10994-006-6889-7. URL https://link.springer. com/article/10.1007/s10994-006-6889-7

    work page doi:10.1007/s10994-006-6889-7

  31. [31]

    Generalized Score Functions for Causal Discovery.KDD : proceedings

    Biwei Huang, Kun Zhang, Yizhu Lin, Bernhard Schölkopf, and Clark Glymour. Generalized Score Functions for Causal Discovery.KDD : proceedings. International Conference on Knowledge Discovery & Data Mining, 2018:1551–1560, 8 2018. ISSN 2154-817X. doi: 10. 1145/3219819.3220104. URL http://www.ncbi.nlm.nih.gov/pubmed/30191079http: //www.pubmedcentral.nih.gov/...

    arXiv 2018

  32. [32]

    DAGs with NO TEARS: Continuous optimization for structure learning

    Xun Zheng, Bryon Aragam, Pradeep K Ravikumar, and Eric P Xing. DAGs with NO TEARS: Continuous optimization for structure learning. InAdvances in Neural Information Processing Systems, volume 31, pages 9472–9483, 2018

    2018

  33. [33]

    Scaling structural learning with no-bears to infer causal transcriptome networks

    Hao-Chih Lee, Matteo Danieletto, Riccardo Miotto, Sarah T Cherng, and Joel T Dudley. Scaling structural learning with no-bears to infer causal transcriptome networks. InPACIFIC SYMPOSIUM ON BIOCOMPUTING 2020, pages 391–402. World Scientific, 2019

    2020

  34. [34]

    DAG-GNN: DAG Structure Learning with Graph Neural Networks

    Yue Yu, Jie Chen, Tian Gao, and Mo Yu. DAG-GNN: DAG Structure Learning with Graph Neural Networks.36th International Conference on Machine Learning, ICML 2019, 2019-June: 12395–12406, 2019. doi: 10.48550/arxiv.1904.10098

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1904.10098 2019

  35. [35]

    Gradient- based neural dag learning.arXiv preprint arXiv:1906.02226, 2019

    Sébastien Lachapelle, Philippe Brouillard, Tristan Deleu, and Simon Lacoste-Julien. Gradient- based neural dag learning.arXiv preprint arXiv:1906.02226, 2019

    arXiv 1906

  36. [36]

    Learning sparse nonparametric dags

    Xun Zheng, Chen Dan, Bryon Aragam, Pradeep Ravikumar, and Eric Xing. Learning sparse nonparametric dags. InInternational Conference on Artificial Intelligence and Statistics, pages 3414–3425. PMLR, 2020

    2020

  37. [37]

    Differentiable causal discovery from interventional data

    Philippe Brouillard, Sébastien Lachapelle, Alexandre Lacoste, Simon Lacoste-Julien, and Alexandre Drouin. Differentiable causal discovery from interventional data. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors,Advances in Neural Information Processing Systems, volume 33, pages 21865–21877. Curran Associates, Inc., 2020. 12

    2020

  38. [38]

    A meta-transfer objective for learning to disentangle causal mechanisms.arXiv preprint arXiv:1901.10912, 2019

    Yoshua Bengio, Tristan Deleu, Nasim Rahaman, Rosemary Ke, Sébastien Lachapelle, Olexa Bilaniuk, Anirudh Goyal, and Christopher Pal. A meta-transfer objective for learning to disentangle causal mechanisms.arXiv preprint arXiv:1901.10912, 2019

    Pith/arXiv arXiv 1901

  39. [39]

    Learning Neural Causal Models from Unknown Interventions

    Nan Rosemary Ke, Olexa Bilaniuk, Anirudh Goyal, Stefan Bauer, Hugo Larochelle, Bernhard Schölkopf, Michael C Mozer, Chris Pal, and Yoshua Bengio. Learning Neural Causal Models from Unknown Interventions

  40. [40]

    Efficient Neural Causal Discovery without Acyclicity Constraints

    Phillip Lippe, Taco Cohen, Qualcomm Ai Research, and Efstratios Gavves. Efficient Neural Causal Discovery without Acyclicity Constraints. 2021. doi: 10.48550/arxiv.2107.10483

    work page doi:10.48550/arxiv.2107.10483 2021

  41. [41]

    Feature Selection: A Data Perspective

    Jundong Li, Kewei Cheng, Suhang Wang, Fred Morstatter, Robert P. Trevino, Jiliang Tang, and Huan Liu. Feature Selection: A Data Perspective.ACM Computing Surveys, 50(6), 1 2016. doi: 10.1145/3136625. URL http://arxiv.org/abs/1601.07996http://dx.doi.org/ 10.1145/3136625

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1145/3136625 2016

  42. [42]

    A Unified View of Causal and Non-causal Feature Selection

    Kui Yu, Lin Liu, and Jiuyong Li. A Unified View of Causal and Non-causal Feature Selection. ACM Transactions on Knowledge Discovery from Data, 15(4), 2 2018. ISSN 1556472X. doi: 10.48550/arxiv.1802.05844. URLhttps://arxiv.org/abs/1802.05844v4

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1802.05844 2018

  43. [43]

    In: Medical Imaging with Deep Learning (MIDL)

    Kui Yu, Xianjie Guo, Lin Liu, Jiuyong Li, Hao Wang, Zhaolong Ling, and Xindong Wu. Causality-based Feature Selection: Methods and Evaluations. 11 2019. doi: 10.48550/arxiv. 1911.07147. URLhttps://arxiv.org/abs/1911.07147

    work page internal anchor Pith review doi:10.48550/arxiv 2019

  44. [44]

    Learning neural causal models from unknown interventions.arXiv preprint arXiv:1910.01075, 2019

    Nan Rosemary Ke, Olexa Bilaniuk, Anirudh Goyal, Stefan Bauer, Hugo Larochelle, Bernhard Schölkopf, Michael C Mozer, Chris Pal, and Yoshua Bengio. Learning neural causal models from unknown interventions.arXiv preprint arXiv:1910.01075, 2019

    arXiv 1910

  45. [45]

    Efficient neural causal discovery without acyclicity constraints.arXiv preprint arXiv:2107.10483, 2021

    Phillip Lippe, Taco Cohen, and Efstratios Gavves. Efficient neural causal discovery without acyclicity constraints.arXiv preprint arXiv:2107.10483, 2021

    arXiv 2021

  46. [46]

    Learning to induce causal structure.arXiv preprint arXiv:2204.04875, 2022

    Nan Rosemary Ke, Silvia Chiappa, Jane Wang, Jorg Bornschein, Theophane Weber, Anirudh Goyal, Matthew Botvinick, Michael Mozer, and Danilo Jimenez Rezende. Learning to Induce Causal Structure. 4 2022. doi: 10.48550/arxiv.2204.04875. URL https://arxiv.org/abs/ 2204.04875v1

    work page doi:10.48550/arxiv.2204.04875 2022

  47. [47]

    Adams, and Nando de Freitas

    Bobak Shahriari, Kevin Swersky, Ziyu Wang, Ryan P. Adams, and Nando de Freitas. Taking the Human Out of the Loop: A Review of Bayesian Optimization.Proceedings of the IEEE, 104(1):148–175, 1 2016. ISSN 0018-9219. doi: 10.1109/JPROC.2015.2494218. URL https: //ieeexplore.ieee.org/document/7352306/

    work page doi:10.1109/jproc.2015.2494218 2016

  48. [48]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization.7th International Conference on Learning Representations, ICLR 2019, 11 2017. doi: 10.48550/arxiv.1711.05101. URLhttps://arxiv.org/abs/1711.05101v3

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1711.05101 2019

  49. [49]

    Bishara and James B

    Anthony J. Bishara and James B. Hittner. Testing the significance of a correlation with nonnormal data: Comparison of Pearson, Spearman, transformation, and resampling approaches. Psychological Methods, 17(3):399–417, 9 2012. ISSN 1082989X. doi: 10.1037/A0028087. URL/record/2012-11642-001

    work page doi:10.1037/a0028087 2012

  50. [50]

    David D. Lewis. Feature Selection and Feature Extraction for Text Categorization. page 212,

  51. [51]

    doi: 10.3115/1075527.1075574

    work page doi:10.3115/1075527.1075574

  52. [52]

    Johnson and Robert E

    Kevin J. Johnson and Robert E. Synovec. Pattern recognition of jet fuels: comprehensive GC×GC with ANOV A-based feature selection and principal component analysis.Chemometrics and Intelligent Laboratory Systems, 60(1-2):225–237, 1 2002. ISSN 0169-7439. doi: 10.1016/ S0169-7439(01)00198-8

    2002

  53. [53]

    Huan Liu and R. Setiono. Chi2: feature selection and discretization of numeric attributes. In Proceedings of 7th IEEE International Conference on Tools with Artificial Intelligence, pages 388–391. IEEE Comput. Soc. Press. ISBN 0-8186-7312-5. doi: 10.1109/TAI.1995.479783. URLhttp://ieeexplore.ieee.org/document/479783/

    work page doi:10.1109/tai.1995.479783 1995

  54. [54]

    Kursa, Aleksander Jankowski, and Witold R

    Miron B. Kursa, Aleksander Jankowski, and Witold R. Rudnicki. Boruta – A System for Feature Selection.Fundamenta Informaticae, 101(4):271–285, 1 2010. ISSN 01692968. doi: 10.3233/FI-2010-288. URL https://www.medra.org/servlet/aliasResolver?alias= iospress&doi=10.3233/FI-2010-288. 13

    work page doi:10.3233/fi-2010-288 2010

  55. [55]

    Swamping and masking in Markov boundary discovery.Machine Learning, 104(1):25–54, 7 2016

    Xuqing Liu and Xinsheng Liu. Swamping and masking in Markov boundary discovery.Machine Learning, 104(1):25–54, 7 2016. ISSN 0885-6125. doi: 10.1007/s10994-016-5545-0. URL http://link.springer.com/10.1007/s10994-016-5545-0

    work page doi:10.1007/s10994-016-5545-0 2016

  56. [56]

    Efficient Markov Blanket Discovery and Its Application.IEEE transactions on cybernetics, 47(5):1169–1179, 5 2017

    Tian Gao and Qiang Ji. Efficient Markov Blanket Discovery and Its Application.IEEE transactions on cybernetics, 47(5):1169–1179, 5 2017. ISSN 2168-2275. doi: 10.1109/TCYB. 2016.2539338. URLhttp://www.ncbi.nlm.nih.gov/pubmed/27046886

    work page doi:10.1109/tcyb 2017

  57. [57]

    HITON: a novel Markov Blanket algorithm for op- timal variable selection.AMIA

    C F Aliferis, I Tsamardinos, and A Statnikov. HITON: a novel Markov Blanket algorithm for op- timal variable selection.AMIA ... Annual Symposium proceedings. AMIA Symposium, 2003:21–5,

    2003

  58. [58]

    URL http://www.ncbi.nlm.nih.gov/pubmed/14728126http: //www.pubmedcentral.nih.gov/articlerender.fcgi?artid=PMC1480117

    ISSN 1942-597X. URL http://www.ncbi.nlm.nih.gov/pubmed/14728126http: //www.pubmedcentral.nih.gov/articlerender.fcgi?artid=PMC1480117

    arXiv 1942

  59. [59]

    Tsamardinos, C

    I. Tsamardinos, C. Aliferis, and A. Statnikov. Algorithms for Large Scale Markov Blanket Discovery. 2003

    2003

  60. [60]

    Learning linear non-Gaussian graphical models with multidirected edges.Journal of Causal Inference, 9(1):250–263, 1 2021

    Yiheng Liu, Elina Robeva, and Huanqing Wang. Learning linear non-Gaussian graphical models with multidirected edges.Journal of Causal Inference, 9(1):250–263, 1 2021. ISSN 21933685. doi: 10.1515/JCI-2020-0027

    work page doi:10.1515/jci-2020-0027 2021

  61. [61]

    URL https://archive.ics.uci.edu/ ml/citation_policy.html

    UCI Machine Learning Repository: Citation Policy. URL https://archive.ics.uci.edu/ ml/citation_policy.html

  62. [62]

    CTAB-GAN: Effective Table Data Synthesizing, 11 2021

    Zilong Zhao, Aditya Kunar, Robert Birke, Lydia Y Chen Lydiaychen, Vineeth N Balasubra- manian, and Ivor Tsang. CTAB-GAN: Effective Table Data Synthesizing, 11 2021. ISSN 2640-3498. URLhttps://proceedings.mlr.press/v157/zhao21a.html

    2021

  63. [63]

    Yu. V . Gorishniy, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Revisiting Deep Learning Models for Tabular Data.ArXiv, 2021

    2021

  64. [64]

    TabDDPM: Modelling Tabular Data with Diffusion Models

    Akim Kotelnikov HSE, Dmitry Baranchuk, Ivan Rubachev HSE, and Yandex Artem Babenko Yandex. TabDDPM: Modelling Tabular Data with Diffusion Models. 2022. doi: 10.48550/arxiv.2209.15421

    work page doi:10.48550/arxiv.2209.15421 2022

  65. [65]

    Mostafizur Rahman and D

    M. Mostafizur Rahman and D. N. Davis. Addressing the Class Imbalance Problem in Medical Datasets.International Journal of Machine Learning and Computing, pages 224–228, 2013. ISSN 20103700. doi: 10.7763/ijmlc.2013.v3.307

    work page doi:10.7763/ijmlc.2013.v3.307 2013

  66. [66]

    P. H. Bennett, N. B. Rushforth, M. Miller, and P. M. LeCompte. Epidemiologic studies of diabetes in the Pima Indians.Recent Progress in Hormone Research, V ol. 32:333–376, 1976. ISSN 00799963. doi: 10.1016/b978-0-12-571132-6.50021-x

    work page doi:10.1016/b978-0-12-571132-6.50021-x 1976

  67. [67]

    Hupfeld and Jerrold M

    Christopher J. Hupfeld and Jerrold M. Olefsky. Type 2 Diabetes Mellitus: Etiology, Patho- genesis, and Natural History.Endocrinology: Adult and Pediatric, 1-2:691–714, 1 2015. doi: 10.1016/B978-0-323-18907-1.00040-8

    work page doi:10.1016/b978-0-323-18907-1.00040-8 2015

  68. [68]

    Larry Jameson, David M

    J. Larry Jameson, David M. de Kretser, Ashley B. Grossman, John T. Potts, J. Larry De Groot, Linda C. Giudice, Shlomo Melmed, and Gordon C. Weir.Endocrinology: Adult and Pediatric, volume 1-2. Elsevier Inc., 1 2015. ISBN 9780323189071. doi: 10.1016/ c2012-1-03052-4. URL http://www.sciencedirect.com:5070/book/9780323189071/ endocrinology-adult-and-pediatric

    arXiv 2015

  69. [69]

    Jeffrey Metter, B

    E. Jeffrey Metter, B. Gwen Windham, Marcello Maggio, Eleanor M. Simonsick, Shari M. Ling, Josephine M. Egan, and Luigi Ferrucci. Glucose and Insulin Measurements from the Oral Glucose Tolerance Test and Mortality Prediction.Diabetes Care, 31(5):1026–1030, 5

  70. [70]

    doi: 10.2337/DC07-2102

    ISSN 0149-5992. doi: 10.2337/DC07-2102. URL https://dx.doi.org/10.2337/ dc07-2102

    work page doi:10.2337/dc07-2102

  71. [71]

    whitegrid

    Lukas Biewald. Experiment Tracking with Weights and Biases.Software available from wandb.com, 2020. URLhttps://www.wandb.com/. 14 Contents 1 Introduction 1 2 Background 3 3 Related Works 3 4 Methodology 4 4.1 Differentiable Causal Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 4.2 Detecting natural experiments . . . . . . . . . . . . . ...

    2020