Machine learning models for estimating counterfactuals in a single-arm inflammatory bowel disease study

arxiv: 2604.23465 · v1 · submitted 2026-04-25 · 💻 cs.LG

Machine learning models for estimating counterfactuals in a single-arm inflammatory bowel disease study

Dan Liu , Fida K. Dankar , Jennifer C. deBruyn , Amanda Ricciuto , Anne M. Griffiths , Thomas D. Walters , Khaled EI Emam This is my paper

Pith reviewed 2026-05-08 08:20 UTC · model grok-4.3

classification 💻 cs.LG

keywords machine learningcounterfactual estimationvirtual controlssingle-arm trialsinflammatory bowel diseasepediatric Crohn's diseasepropensity score matching

0 comments p. Extension

The pith

Machine learning models trained on external IFX data can generate virtual control arms yielding treatment effect estimates that match propensity score matching in a single-arm pediatric Crohn's study.

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

This paper develops and evaluates machine learning models to create virtual control arms for single-arm trials by training on external data from IFX-treated patients. The models predict counterfactual outcomes of steroid-free clinical remission and a composite CRP-SFCR measure for ADA-treated pediatric Crohn's patients. LightGBM produces the odds ratio closest to the reference from propensity score matching. All resulting 95% confidence intervals support the same conclusion as the matched analysis: no statistical difference between the treatments. The work shows that such virtual controls can substitute for recruiting real concurrent control patients in IBD studies.

Core claim

Training five machine learning models on an external IFX-treated cohort to predict 1-year SFCR and CRP-SFCR counterfactual outcomes for ADA-treated patients produces odds ratios and confidence intervals that closely match those from propensity score matching to external controls, with LightGBM performing best and all intervals consistent with no statistical difference between ADA and IFX.

What carries the argument

Counterfactual prediction models (especially LightGBM) trained on IFX external cohort data to simulate control-arm responses for the ADA single-arm cohort.

Load-bearing premise

The IFX-treated external cohort provides a sufficiently representative basis for training models that accurately predict counterfactual outcomes in the ADA-treated cohort without major unmeasured confounding or distribution shift.

What would settle it

A randomized trial directly comparing ADA and IFX in a similar pediatric Crohn's population that finds a statistically significant difference in SFCR or CRP-SFCR would contradict the virtual control estimates.

read the original abstract

Single-arm trials accelerate study timelines by reducing the number of patients that must be recruited for a concurrent control group. However, these designs require an alternative comparator to estimate treatment effects. One approach is to construct a virtual control arm using a machine learning (ML) model trained on external control data to predict the counterfactual outcomes of the treatment arm. Our aim in this study was to leverage virtual controls by developing and evaluating ML-based counterfactual outcome models trained on IFX-treated patients to predict 1-year steroid-free clinical remission (SFCR ) and a composite of C-reactive protein remission plus steroid-free clinical remission (CRP-SFCR) for ADA-treated pediatric Crohn's disease patients, and to compare the resulting IFX-versus-ADA treatment effect estimates with those obtained using propensity score matching to external controls. Five ML models were used to train counterfactual models on the observed IFX cohort data. The resulting models were used to predict the counterfactual outcomes for the ADA arm patients. LGBM yields the best OR closest to the propensity score matched reference, and all 95% CI results align with the conclusion from the reference study that no statistical difference in the primary and secondary outcomes has been observed between the patients treated with ADA or IFX. Our study supports virtual controls as a viable and effective substitute for expensive, lengthy or unethical patient recruitment in an inflammatory bowel disease (IBD) trial. The developed gradient boosted prediction model can be used as a pretrained model to generate IFX counterfactual predictions in future studies, pending external validation and assessment of transportability.

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.

Desk Editor's Note private letter to a colleague

The paper applies standard ML counterfactual models to an IFX-to-ADA comparison in pediatric Crohn's and gets numerical agreement with PSM, but leaves transportability untested.

read the letter

The main takeaway is that LGBM trained on an external IFX cohort produces odds ratios and confidence intervals for steroid-free remission that sit close to the propensity-score-matched reference when applied to the ADA arm. That alignment is the concrete result, and it is useful to see five different models run head-to-head on the same task rather than just one chosen method. The work also flags the pretrained LGBM as something that could be reused in later studies, which is a practical step if the transportability claim holds up later. What is new here is the specific head-to-head in this pediatric IBD setting with direct numerical comparison to PSM; it is not a new algorithm but a targeted application that shows the numbers can line up. The paper does a clean job of stating the clinical conclusion (no detectable difference between ADA and IFX) and showing that the ML route reaches the same place. The soft spot is the transportability assumption. The IFX training data and ADA target cohort come from different treatment groups, and nothing in the reported work checks covariate overlap, propensity-score distributions, or sensitivity to possible shifts in unmeasured factors. Without those diagnostics the agreement with PSM could reflect shared bias rather than valid counterfactuals. Model training and validation details are also thin in the abstract, so it is hard to judge how much the result depends on particular feature choices or hyperparameter decisions. This is the kind of paper that belongs in a methods or applied statistics journal for trialists who already work with external controls. Readers who need a worked example of ML virtual arms in IBD will get value from the comparison, even if they will want the full methods to assess robustness. It is worth sending to peer review because the core idea is straightforward, the comparison is direct, and the clinical setting is one where reducing control-arm size matters. A referee can ask for the missing overlap checks and training protocol without the paper needing to be rewritten from scratch.

Referee Report

3 major / 2 minor

Summary. The paper develops and evaluates five machine learning models (including LGBM) trained on an external IFX-treated cohort to predict counterfactual 1-year steroid-free clinical remission (SFCR) and CRP-SFCR outcomes for ADA-treated pediatric Crohn's disease patients in a single-arm study. The resulting treatment effect estimates (odds ratios and 95% CIs) are compared to a propensity score matching (PSM) reference analysis on external controls; LGBM produces the OR closest to the PSM benchmark, and all model-derived CIs are consistent with the reference conclusion of no statistically significant difference between ADA and IFX. The work concludes that virtual controls via ML are a viable substitute for concurrent controls in IBD trials and offers the LGBM model as a pretrained tool for future use, pending external validation.

Significance. If the transportability and calibration claims hold after proper validation, the approach could meaningfully accelerate single-arm IBD trials by substituting external-data-driven counterfactuals for randomized controls, with direct implications for pediatric Crohn's studies where ethical or recruitment barriers are high. The explicit benchmarking against PSM and the offer of a reusable pretrained model are concrete strengths that would support reproducibility if accompanied by code, diagnostics, and sensitivity checks.

major comments (3)

[Abstract] Abstract and Methods (model training and application): No details are provided on model training procedures, validation (e.g., cross-validation strategy, hold-out sets), hyperparameter tuning, feature preprocessing, handling of missing data, or performance metrics (AUC, calibration, Brier score) on the IFX training data. Without these, the claim that LGBM produces the 'best' OR cannot be evaluated and the numerical agreement with PSM may reflect shared bias rather than valid counterfactual estimation.
[Abstract] Abstract and Results (counterfactual prediction): The manuscript reports no diagnostics for covariate overlap, propensity-score distribution overlap, or transportability between the IFX training cohort and ADA target cohort (e.g., no overlap plots, standardized mean differences, or sensitivity analyses for unmeasured confounding or distribution shift). This assumption is load-bearing for the headline result that LGBM ORs and CIs align with the PSM reference.
[Abstract] Abstract: The statement that 'all 95% CI results align with the conclusion... no statistical difference' is presented without the actual OR/CI values, sample sizes, or variance estimation method for the ML-derived estimates, making it impossible to assess whether the alignment is substantive or merely non-informative.

minor comments (2)

[Abstract] The abstract refers to 'five ML models' but does not name them beyond LGBM; listing the full set (e.g., in a table) would improve clarity.
[Abstract] Clarify whether the PSM reference uses the same external IFX cohort or a distinct one, and whether any patients overlap between training and target cohorts.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important areas for improving the transparency and rigor of our work on ML-based virtual controls in single-arm IBD studies. We address each major comment point by point below and have prepared revisions to add the requested details, diagnostics, and numerical results.

read point-by-point responses

Referee: [Abstract] Abstract and Methods (model training and application): No details are provided on model training procedures, validation (e.g., cross-validation strategy, hold-out sets), hyperparameter tuning, feature preprocessing, handling of missing data, or performance metrics (AUC, calibration, Brier score) on the IFX training data. Without these, the claim that LGBM produces the 'best' OR cannot be evaluated and the numerical agreement with PSM may reflect shared bias rather than valid counterfactual estimation.

Authors: We agree that the abstract and Methods section as currently written do not provide these implementation details, which are necessary to fully evaluate model selection and the risk of shared bias with PSM. The manuscript describes training five models (including LGBM) on the external IFX cohort but omits the requested specifics. In the revised version we will expand the Methods with a dedicated model development subsection reporting: stratified 5-fold cross-validation on the IFX data, grid-search hyperparameter tuning, feature preprocessing steps, missing-data handling via multiple imputation, and performance metrics (AUC, calibration plots, Brier scores) for all five models. These additions will allow readers to assess whether LGBM's closer alignment with the PSM benchmark reflects genuine counterfactual validity. revision: yes
Referee: [Abstract] Abstract and Results (counterfactual prediction): The manuscript reports no diagnostics for covariate overlap, propensity-score distribution overlap, or transportability between the IFX training cohort and ADA target cohort (e.g., no overlap plots, standardized mean differences, or sensitivity analyses for unmeasured confounding or distribution shift). This assumption is load-bearing for the headline result that LGBM ORs and CIs align with the PSM reference.

Authors: This is a substantive point; the transportability assumption is indeed central and was not explicitly diagnosed in the current manuscript. While the PSM benchmark implicitly relies on overlap, we did not report supporting diagnostics. The revised manuscript will add: covariate overlap plots, standardized mean differences between IFX and ADA cohorts, propensity-score distribution comparisons, and sensitivity analyses for unmeasured confounding and distribution shift (including simulation-based checks). These will be presented alongside the main results to strengthen the justification for applying the IFX-trained model to the ADA cohort. revision: yes
Referee: [Abstract] Abstract: The statement that 'all 95% CI results align with the conclusion... no statistical difference' is presented without the actual OR/CI values, sample sizes, or variance estimation method for the ML-derived estimates, making it impossible to assess whether the alignment is substantive or merely non-informative.

Authors: We acknowledge that the abstract's qualitative summary of alignment lacks the quantitative detail needed for evaluation. The manuscript states that LGBM yields the OR closest to PSM and that all CIs are consistent with no difference, but does not report the numbers. In the revision we will update the abstract and Results to include the specific ORs and 95% CIs from the LGBM model and the PSM reference, the sample sizes for the IFX training cohort and ADA target cohort, and the variance estimation method (bootstrap resampling of the predicted counterfactual outcomes). This will make the degree of alignment transparent and evaluable. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external training data and independent benchmark

full rationale

The paper trains five ML models (including LGBM) exclusively on the observed IFX-treated external cohort to predict counterfactual SFCR and CRP-SFCR outcomes for the ADA-treated target cohort using their observed covariates. The resulting ORs and 95% CIs are then compared to a separate propensity-score-matched analysis performed on external controls. No equations, fitting procedures, or self-citations reduce the reported treatment-effect estimates to parameters defined by the ADA target data itself. The method is self-contained against the external IFX training set and the independent PSM reference; transportability assumptions are stated but do not create definitional or fitted-input circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that observational IFX data can be used to train models that generalize to counterfactual prediction for ADA patients; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)

domain assumption The IFX cohort data is representative and transportable to the ADA population for counterfactual prediction.
Models trained on IFX data are applied directly to ADA patients; transportability is invoked without detailed justification or sensitivity checks in the abstract.

pith-pipeline@v0.9.0 · 5602 in / 1325 out tokens · 26298 ms · 2026-05-08T08:20:36.566924+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

15 extracted references · 13 canonical work pages

[1]

Randomized controlled trials – The what, when, how and why

1 Braga LH, Farrokhyar F, Dönmez Mİ, et al. Randomized controlled trials – The what, when, how and why. Journal of Pediatric Urology. 2025;21:397–404. doi: 10.1016/j.jpurol.2024.11.021 2 Griessbach A, Speich B, Amstutz A, et al. Resource use and costs of investigator-sponsored randomized clinical trials in Switzerland, Germany, and the United Kingdom: a m...

work page doi:10.1016/j.jpurol.2024.11.021 2025
[2]

Cost-effectiveness of health research study participant recruitment strategies: a systematic review

19/21 6 Huynh L, Johns B, Liu S-H, et al. Cost-effectiveness of health research study participant recruitment strategies: a systematic review. Clin Trials. 2014;11:576–83. doi: 10.1177/1740774514540371 7 Lambert J, Lengliné E, Porcher R, et al. Enriching single-arm clinical trials with external controls: possibilities and pitfalls. Blood Adv. 2023;7:5680–...

work page doi:10.1177/1740774514540371 2014
[3]

https://arxiv.org/abs/2405.12437v2 (accessed 31 March

work page arXiv
[4]

Increasing FDA-accelerated approval of single-arm trials in oncology (1992 to 2020)

17 Ribeiro TB, Bennett CL, Colunga-Lozano LE, et al. Increasing FDA-accelerated approval of single-arm trials in oncology (1992 to 2020). Journal of Clinical Epidemiology. 2023;159:151–8. doi: 10.1016/j.jclinepi.2023.04.001 18 Nuño MM, Pugh SL, Ji L, et al. On the use of external controls in clinical trials. J Natl Cancer Inst Monogr. 2025;2025:30–4. doi:...

work page doi:10.1016/j.jclinepi.2023.04.001 1992
[5]

Phenotypic Variation in Paediatric Inflammatory Bowel Disease by Age: A Multicentre Prospective Inception Cohort Study of the Canadian Children IBD Network

34 Dhaliwal J, Walters TD, Mack DR, et al. Phenotypic Variation in Paediatric Inflammatory Bowel Disease by Age: A Multicentre Prospective Inception Cohort Study of the Canadian Children IBD Network. J Crohns Colitis. 2020;14:445–54. doi: 10.1093/ecco-jcc/jjz106 35 Fernandes A, Porcher R, Tran V-T, et al. Evaluating virtual-control-augmented trials for re...

work page doi:10.1093/ecco-jcc/jjz106 2020
[6]

Hollmann, S

36 Hollmann N, Müller S, Purucker L, et al. Accurate predictions on small data with a tabular foundation model. Nature. 2025;637:319–26. doi: 10.1038/s41586-024-08328-6 37 Stone M. Cross-Validatory Choice and Assessment of Statistical Predictions. Journal of the Royal Statistical Society Series B (Methodological). 1974;36:111–47. 38 Varma S, Simon R. Bias...

work page doi:10.1038/s41586-024-08328-6 2025
[7]

https://arxiv.org/abs/2107.05847v3 (accessed 9 December

work page arXiv
[8]

https://arxiv.org/abs/1806.10961v3 (accessed 9 December

work page arXiv
[9]

Performance evaluation of predictive AI models to support medical decisions: Overview and guidance

47 Calster BV, Collins GS, Vickers AJ, et al. Performance evaluation of predictive AI models to support medical decisions: Overview and guidance. 48 Richardson E, Trevizani R, Greenbaum JA, et al. The receiver operating characteristic curve accurately assesses imbalanced datasets. Patterns. 2024;5:100994. doi: 10.1016/j.patter.2024.100994 49 Brabec J, Kom...

work page doi:10.1016/j.patter.2024.100994 2024
[10]

doi: 10.1007/978-3-030-50423-6_6 50 Austin PC, Steyerberg EW

2020;12140:74–87. doi: 10.1007/978-3-030-50423-6_6 50 Austin PC, Steyerberg EW. The Integrated Calibration Index (ICI) and related metrics for quantifying the calibration of logistic regression models. Stat Med. 2019;38:4051–65. doi: 10.1002/sim.8281 51 Watson DS, Blesch K, Kapar J, et al. Adversarial random forests for density estimation and generative modeling

work page doi:10.1007/978-3-030-50423-6_6 2020
[11]

pgmpy: Probabilistic Graphical Models using Python

52 Ankan A, Panda A. pgmpy: Probabilistic Graphical Models using Python. Austin, Texas 2015:6–11. 53 Xu L, Skoularidou M, Cuesta-Infante A, et al. Modeling Tabular data using Conditional GAN. Advances in Neural Information Processing Systems

2015
[12]

Synthcity: facilitating innovative use cases of synthetic data in different data modalities

54 Qian Z, Cebere B-C, van der Schaar M. Synthcity: facilitating innovative use cases of synthetic data in different data modalities. arXiv. 2023;2301.07573. 55 Goring S, Taylor A, Müller K, et al. Characteristics of non-randomised studies using comparisons with external controls submitted for regulatory approval in the USA and Europe: a systematic review...

work page doi:10.1136/bmjopen-2018-024895 2023
[13]

Application of Bayesian networks to generate synthetic health data

2 Kaur D, Sobiesk M, Patil S, et al. Application of Bayesian networks to generate synthetic health data. J Am Med Inform Assoc. 2021;28:801–11. doi: 10.1093/jamia/ocaa303 3 Murphy KP. Machine Learning: A Probabilistic Perspective. MIT Press

work page doi:10.1093/jamia/ocaa303 2021
[14]

2014:2672–80

Montreal, QC, Canada: Neural Information Processing Systems Foundation, Inc. 2014:2672–80. 5 Bourou S, El Saer A, Velivassaki T-H, et al. A Review of Tabular Data Synthesis Using GANs on an IDS Dataset. Information. 2021;12:375. doi: 10.3390/info12090375 6 Xu L, Skoularidou M, Cuesta-Infante A, et al. Modeling Tabular data using Conditional GAN. Advances ...

work page doi:10.3390/info12090375 2014
[15]

Variational autoencoder based synthetic data generation for imbalanced learning

8 Wan Z, Zhang Y, He H. Variational autoencoder based synthetic data generation for imbalanced learning. 2017 IEEE Symposium Series on Computational Intelligence (SSCI). 2017:1–7. 9 Ishfaq H, Hoogi A, Rubin D. TVAE: Triplet-Based Variational Autoencoder using Metric Learning

2017

[1] [1]

Randomized controlled trials – The what, when, how and why

1 Braga LH, Farrokhyar F, Dönmez Mİ, et al. Randomized controlled trials – The what, when, how and why. Journal of Pediatric Urology. 2025;21:397–404. doi: 10.1016/j.jpurol.2024.11.021 2 Griessbach A, Speich B, Amstutz A, et al. Resource use and costs of investigator-sponsored randomized clinical trials in Switzerland, Germany, and the United Kingdom: a m...

work page doi:10.1016/j.jpurol.2024.11.021 2025

[2] [2]

Cost-effectiveness of health research study participant recruitment strategies: a systematic review

19/21 6 Huynh L, Johns B, Liu S-H, et al. Cost-effectiveness of health research study participant recruitment strategies: a systematic review. Clin Trials. 2014;11:576–83. doi: 10.1177/1740774514540371 7 Lambert J, Lengliné E, Porcher R, et al. Enriching single-arm clinical trials with external controls: possibilities and pitfalls. Blood Adv. 2023;7:5680–...

work page doi:10.1177/1740774514540371 2014

[3] [3]

https://arxiv.org/abs/2405.12437v2 (accessed 31 March

work page arXiv

[4] [4]

Increasing FDA-accelerated approval of single-arm trials in oncology (1992 to 2020)

17 Ribeiro TB, Bennett CL, Colunga-Lozano LE, et al. Increasing FDA-accelerated approval of single-arm trials in oncology (1992 to 2020). Journal of Clinical Epidemiology. 2023;159:151–8. doi: 10.1016/j.jclinepi.2023.04.001 18 Nuño MM, Pugh SL, Ji L, et al. On the use of external controls in clinical trials. J Natl Cancer Inst Monogr. 2025;2025:30–4. doi:...

work page doi:10.1016/j.jclinepi.2023.04.001 1992

[5] [5]

Phenotypic Variation in Paediatric Inflammatory Bowel Disease by Age: A Multicentre Prospective Inception Cohort Study of the Canadian Children IBD Network

34 Dhaliwal J, Walters TD, Mack DR, et al. Phenotypic Variation in Paediatric Inflammatory Bowel Disease by Age: A Multicentre Prospective Inception Cohort Study of the Canadian Children IBD Network. J Crohns Colitis. 2020;14:445–54. doi: 10.1093/ecco-jcc/jjz106 35 Fernandes A, Porcher R, Tran V-T, et al. Evaluating virtual-control-augmented trials for re...

work page doi:10.1093/ecco-jcc/jjz106 2020

[6] [6]

Hollmann, S

36 Hollmann N, Müller S, Purucker L, et al. Accurate predictions on small data with a tabular foundation model. Nature. 2025;637:319–26. doi: 10.1038/s41586-024-08328-6 37 Stone M. Cross-Validatory Choice and Assessment of Statistical Predictions. Journal of the Royal Statistical Society Series B (Methodological). 1974;36:111–47. 38 Varma S, Simon R. Bias...

work page doi:10.1038/s41586-024-08328-6 2025

[7] [7]

https://arxiv.org/abs/2107.05847v3 (accessed 9 December

work page arXiv

[8] [8]

https://arxiv.org/abs/1806.10961v3 (accessed 9 December

work page arXiv

[9] [9]

Performance evaluation of predictive AI models to support medical decisions: Overview and guidance

47 Calster BV, Collins GS, Vickers AJ, et al. Performance evaluation of predictive AI models to support medical decisions: Overview and guidance. 48 Richardson E, Trevizani R, Greenbaum JA, et al. The receiver operating characteristic curve accurately assesses imbalanced datasets. Patterns. 2024;5:100994. doi: 10.1016/j.patter.2024.100994 49 Brabec J, Kom...

work page doi:10.1016/j.patter.2024.100994 2024

[10] [10]

doi: 10.1007/978-3-030-50423-6_6 50 Austin PC, Steyerberg EW

2020;12140:74–87. doi: 10.1007/978-3-030-50423-6_6 50 Austin PC, Steyerberg EW. The Integrated Calibration Index (ICI) and related metrics for quantifying the calibration of logistic regression models. Stat Med. 2019;38:4051–65. doi: 10.1002/sim.8281 51 Watson DS, Blesch K, Kapar J, et al. Adversarial random forests for density estimation and generative modeling

work page doi:10.1007/978-3-030-50423-6_6 2020

[11] [11]

pgmpy: Probabilistic Graphical Models using Python

52 Ankan A, Panda A. pgmpy: Probabilistic Graphical Models using Python. Austin, Texas 2015:6–11. 53 Xu L, Skoularidou M, Cuesta-Infante A, et al. Modeling Tabular data using Conditional GAN. Advances in Neural Information Processing Systems

2015

[12] [12]

Synthcity: facilitating innovative use cases of synthetic data in different data modalities

54 Qian Z, Cebere B-C, van der Schaar M. Synthcity: facilitating innovative use cases of synthetic data in different data modalities. arXiv. 2023;2301.07573. 55 Goring S, Taylor A, Müller K, et al. Characteristics of non-randomised studies using comparisons with external controls submitted for regulatory approval in the USA and Europe: a systematic review...

work page doi:10.1136/bmjopen-2018-024895 2023

[13] [13]

Application of Bayesian networks to generate synthetic health data

2 Kaur D, Sobiesk M, Patil S, et al. Application of Bayesian networks to generate synthetic health data. J Am Med Inform Assoc. 2021;28:801–11. doi: 10.1093/jamia/ocaa303 3 Murphy KP. Machine Learning: A Probabilistic Perspective. MIT Press

work page doi:10.1093/jamia/ocaa303 2021

[14] [14]

2014:2672–80

Montreal, QC, Canada: Neural Information Processing Systems Foundation, Inc. 2014:2672–80. 5 Bourou S, El Saer A, Velivassaki T-H, et al. A Review of Tabular Data Synthesis Using GANs on an IDS Dataset. Information. 2021;12:375. doi: 10.3390/info12090375 6 Xu L, Skoularidou M, Cuesta-Infante A, et al. Modeling Tabular data using Conditional GAN. Advances ...

work page doi:10.3390/info12090375 2014

[15] [15]

Variational autoencoder based synthetic data generation for imbalanced learning

8 Wan Z, Zhang Y, He H. Variational autoencoder based synthetic data generation for imbalanced learning. 2017 IEEE Symposium Series on Computational Intelligence (SSCI). 2017:1–7. 9 Ishfaq H, Hoogi A, Rubin D. TVAE: Triplet-Based Variational Autoencoder using Metric Learning

2017