Pith. sign in

REVIEW 5 major objections 6 minor 3 cited by

MIRRAMS: Learning Robust Tabular Models under Unseen Missingness Shifts

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read MIRRAMS is a deep-learning objective whose three cross-entropy losses train tabular models that stay accurate when missingness patterns shift from training to test, without assuming how values go missing.

desk verdict The empirical objective is a credible, useful drop-in for tabular learning under missingness shifts, but the support-coverage argument for the theoretical guarantee is false as stated. read the letter →

arxiv 2507.08280 v2 pith:5PCBH2Y4 submitted 2025-07-11 stat.ML cs.LG

Jihye Lee , Minseo Kang , Dongha Kim This is my paper
classification stat.MLcs.LG
keywords missingnessshifttabulardatamissingmutualinformationdistributionsemi-supervisedlearningconsistencyregularizationrobustprediction
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

This paper tries to establish that a single training objective, MIRRAMS, can make tabular prediction models robust to shifts in the missingness pattern between training and test time, even when the test distribution is unseen and no MCAR, MAR, or MNAR assumption is made. The objective is built from three cross-entropy losses: the supervised loss on observed inputs, the same loss on inputs given extra Bernoulli masking, and a high-confidence consistency loss that asks the masked view to reproduce the original view's pseudo-label. The authors argue that these losses enforce mutual-information robustness conditions, so the model keeps label-relevant information and discards pattern-specific noise. Across ten benchmark datasets, MIRRAMS reports the highest average test AUC in every setting shown in Table 1, including the fully observed case, and its semi-supervised variant also does well when training labels are partly missing. A by-product of the derivation is a mutual-information interpretation of FixMatch-style consistency regularization.

What carries the argument

Two objects carry the argument. The first is the additional masking operator $M_r^m$, a product of independent Bernoulli masks that keeps each feature with probability $r$; composing it with the training missingness operator $M_{\mathrm{tr}}$ creates the surrogate distribution $M_r^m \circ M_{\mathrm{tr}} \circ P$ that stands in for the unseen test distribution. The second is the triple of losses derived from conditional-entropy surrogates for mutual information, with Proposition 1 as the bridge: if the model's predictive distribution $P(Y\mid f(\tilde{x}))$ is close in KL divergence to the true $P(Y\mid \tilde{x})$, then the mutual information $I(\tilde{X};Y)$ is close to $I(f(\tilde{X});Y)$. Minimizing $L_1^{\mathrm{MI}}$ and $L_2^{\mathrm{MI}}$ implements that closeness on the observed and masked views, while $L_3^{\mathrm{MI}}$ enforces the consistency side of the conditions. The final objective is $L = L_1^{\mathrm{MI}} + \lambda_1 L_2^{\mathrm{MI}} + \lambda_2 L_3^{\mathrm{MI}}$.

What would settle it

Pick a tabular dataset and declare one feature to be observed in every training example, so $M_{\mathrm{tr}}$ never produces a pattern with that feature missing. Let the test-time operator $M_{\mathrm{ts}}$ mask exactly that feature for a large fraction of examples. Because $M_r^m \circ M_{\mathrm{tr}}$ inherits the constraint that the feature is always observed, the claimed support-coverage statement fails in this construction; observing whether MIRRAMS's accuracy holds or collapses in this setting would settle how much of the robustness depends on the surrogate covering the true test support.

Watch

Extended reading notes

Core claim

The paper's central claim is that satisfying three mutual-information equalities--between the label and the information in the training-missingness view, between the label and the surrogate extra-masked view, and between the two views' predictions--is enough to make a classifier robust to unseen missingness shifts, and that these equalities can be enforced by simple cross-entropy terms. Formally, the modified MI conditions in Eq. (2) are the target; minimizing $L_1^{\mathrm{MI}}$ (supervised cross-entropy on observed inputs), $L_2^{\mathrm{MI}}$ (supervised cross-entropy after extra masking), and $L_3^{\mathrm{MI}}$ (pseudo-label consistency on high-confidence inputs) is presented as a tractable surrogate for enforcing them. The paper's empirical claim is that this objective outperforms eight baselines across MAR, MNAR, and MCAR shifts on ten datasets and remains strong when no missingness is present.

Load-bearing premise

The load-bearing premise is that applying extra Bernoulli masking to the training data creates a surrogate distribution that covers every missingness pattern the test set can produce, so that satisfying the surrogate conditions transfers to the real test distribution.

Editorial extensions

If this is right

  • Under MAR, when training has no missingness and test missingness is 30%, MIRRAMS reports 87.58 average test AUC against 75.33 for XGBoost in Table 1.
  • At the fully observed setting $(\alpha_{\mathrm{tr}},\alpha_{\mathrm{ts}})=(0,0)$, MIRRAMS obtains 91.88 versus 91.17 for XGBoost, so the objective can be used as a general tabular learner even without missingness.
  • With partially missing training labels, the same three-loss framework applied to the labeled and unlabeled subsets outperforms SAINT and SwitchTab across the tested input-output missingness combinations.
  • The loss derivation gives a formal interpretation of FixMatch: its consistency term enforces label-information preservation under augmentation, which offers a mutual-information justification for that class of semi-supervised methods.
  • Because the method makes no MCAR, MAR, or MNAR assumption, the same training procedure applies unchanged across missingness mechanisms.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension: because $L_3^{\mathrm{MI}}$ is structurally identical to FixMatch's consistency loss, the same MI-condition argument should apply to any label-preserving augmentation, not just Bernoulli masking; semantic feature-dropout or learned corruption could be dropped into MIRRAMS with little change.
  • For deployment, the surrogate-coverage premise can be checked before training: if the set of features that are never missing in training is nonempty, the Bernoulli-masked surrogate cannot generate patterns that omit those features, so a test set that concentrates missingness there lies outside the stated guarantee.
  • The fully observed improvement suggests the masked consistency term acts as implicit regularization; one could test this directly by ablating $\lambda_1$ and $\lambda_2$ at $(\alpha_{\mathrm{tr}},\alpha_{\mathrm{ts}})=(0,0)$ and measuring validation-gap behavior rather than test AUC.
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

5 major / 6 minor

Summary. The paper proposes MIRRAMS, a deep learning objective for tabular classification that aims to remain accurate under shifts in the missingness distribution between training and test time. The method defines mutual-information-based robustness conditions, approximates the unseen test missingness distribution by applying an additional Bernoulli masking operator to the training data, and introduces three cross-entropy loss terms (L1, L2, L3) intended to enforce those conditions. Experiments on ten benchmark datasets under MCAR, MAR, and MNAR mechanisms with several (α_tr, α_ts) combinations show that MIRRAMS attains the highest average test AUC in every column of Table 1, including the fully observed setting, and an ablation study reports that both regularization terms contribute. The paper also extends the method to settings where training labels are partially missing, and it offers an interpretation of FixMatch-style consistency regularization in terms of mutual information.

Significance. If the empirical results are taken at face value, MIRRAMS is a simple, efficient, and broadly applicable objective for tabular learning under missingness shifts, with the practically attractive property of improving even fully observed settings. The paper's strengths include a comprehensive experiment suite across ten datasets and three missingness mechanisms, ablations over the main hyperparameters with a fixed default configuration, and a plausible conceptual link between its third loss term and consistency-regularization-based semi-supervised learning. The aggregate result in Table 1—where Ours is highest in all 17 averaged columns—is a meaningful empirical contribution. However, the theoretical claims connecting the surrogate masking distribution to arbitrary test missingness, and the derivation of L3 in particular, are not established as stated, and the 'consistently outperforms' claims are contradicted by several per-dataset tables. The contribution would be better framed as a robust heuristic with strong aggregate empirical support rather than a theoretically guaranteed framework.

major comments (5)
  1. [Proposed Method, MI robustness section, Eq. (2)]
  2. [Derivation of Loss Function in MIRRAMS, Condition (iii), Eqs. (7)-(8)]
  3. [Proposition 1 and Eq. (4)]
  4. [Experiments, Table 1 and Appendix B]
  5. [Experiments, baseline selection]
minor comments (6)
  1. [Notation, before Eq. (8)]
  2. [Notation, Proposed Method]
  3. [Ablation Studies, Confidence Threshold Selection]
  4. [Appendix B, Table B.29]
  5. [Appendix C, Tables 35-38]
  6. [Implementation Details]

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the derivation that L3 enforces MI condition (iii) assumes the model already satisfies the target condition; the empirical evaluation itself is self-contained and not circular.

  1. self definitional [Proposed Method, Derivation of Loss Function, Condition (iii), Eqs. (6)-(8)]
    "Suppose that the training procedure has sufficiently progressed such that f(M_tr(x)) captures nearly all the information relevant to its corresponding label, y. That implies f(M_tr(x)) becomes nearly a one-hot encoding y ... condition (iii) in (2) reduces to the following: H(Y|f(M_tr(X)))≈H(ŷ(M_tr(X))|f(M_m_r∘M_tr(X))). (7)"

    The abstract asserts the losses 'collectively define' an objective that extracts label-relevant information 'regardless of missingness patterns,' and the intro claims each term 'ensures satisfaction of the corresponding component of the MI robustness conditions.' For L3, the chain is: condition (iii) reduces to Eq. (7) only under the supposition that the model has 'sufficiently progressed' so that f(M_tr(x)) is nearly a one-hot encoding of y; L3 is then defined to minimize the RHS of Eq. (7). The target condition is thus presupposed in the derivation of the loss claimed to enforce it, and the pseudo-label target is generated by the same model being trained. The paper concedes: minimizing (8) 'holds only when the overall predictions are made with high confidence.

full rationale

The paper's empirical core is self-contained: Table 1 and the appendix tables report held-out test AUC on ten public benchmarks against baseline implementations, hyperparameters are fixed at (r, λ1, λ2, τ) = (0.2, 15, 15, 0.95) across settings, and no reported quantity is fitted and then re-reported as a prediction. The only self-citation (Kim, Kwon, and Kim 2023) supports a motivational remark and is not load-bearing. The single genuine circular step is in the theoretical derivation of L3: the reduction of condition (iii) to Eq. (7), which licenses L3 (Eq. 8) as enforcing that condition, is obtained under the assumption that 'the training procedure has sufficiently progressed such that f(M_tr(x)) captures nearly all the information relevant to its corresponding label' — i.e., the model already satisfies the essence of the MI conditions the loss is claimed to establish. The paper explicitly concedes that the argument holds only when predictions are made with high confidence, which is consistent with the assumptions under which Eq. (7) is derived. Hence the headline claim that minimizing each term ensures satisfaction of the corresponding MI condition is overstated for L3; the guarantee is conditional on the trained regime it is meant to produce. Separately, the support-coverage argument ('as long as the operator M_tr is non-deterministic, the support of M_m_r∘M_tr∘P always covers that of M_ts∘P for any r > 0') is mathematically false, since composition with Bernoulli masking cannot restore a feature that is always missing under M_tr; however, a false bridge is a correctness risk, not a circular reduction, and does not by itself inflate the circularity score. Overall: partial, disclosed circularity in one theoretical link, with independent empirical content.

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

The central claim rests on five hand-set hyperparameters and per-dataset architecture choices, none of which are derived from theory; several free parameters (r, λ1, λ2) directly control the surrogate distribution and the strength of the regularization. The two load-bearing axioms are the surrogate approximation and the near-one-hot step in the L3 derivation, both appearing in the paper's own text. No invented entities are introduced; the MI robustness conditions are a criterion on the model rather than a new object, and the empirical protocol does not fit test labels, which keeps the circularity burden moderate rather than high.

free parameters (5)
  • additional masking ratio r = 0.2
    Constructs the surrogate distribution M_m_r; the paper says selecting a good r is 'critical' and validates choices by ablation on three datasets, not by theory.
  • loss weight lambda_1 (for L2) = 15
    Selected via ablation on adult, htru2, and qsar bio; fixed across the main experiments.
  • loss weight lambda_2 (for L3) = 15
    Same ablation-based selection; the two weights are not derived from the MI conditions.
  • confidence threshold tau = 0.95
    Gates which pseudo-labels enter L3; ablations show insensitivity in a reasonable range, but the value is chosen by hand.
  • per-dataset architecture sizes = embedding dim 32 vs 4; encoder depth 6 vs 1 (arrhythmia) vs 4 (arcene); heads 8 vs 1; batch 256 vs 64
    Hand-chosen in Appendix B-III for high-dimensional or small datasets; directly affects the reported AUCs with no controlled study of these choices.
assumptions (5)
  • domain assumption The complete joint distribution P of (X, Y) is identical for training and test; only the missingness operators differ.
    Stated in the Notations section; this is the paper's definition of a missingness shift.
  • ad hoc to paper The surrogate distribution M_m_r ∘ M_tr ∘ P approximates the test distribution M_ts ∘ P, so that satisfying Eq. (2) implies satisfying Eq. (1).
    Invoked in the MI robustness conditions section; the supporting support-coverage claim is false in general, so the approximation is an unproven premise.
  • ad hoc to paper Once training has progressed, f(M_tr(X)) is nearly one-hot, so I(f(M_tr(X)); f(M_masked(X))) can be replaced by H(Y) - H(ŷ(M_tr(X)) | f(M_masked(X))).
    Used in the derivation of L3 under condition (iii); it assumes the model already extracts label-relevant information, which is the conclusion the loss is meant to enforce.
  • domain assumption Minimizing empirical cross-entropy drives D_KL(P(Y|x) || P(Y|f(x))) small, so Proposition 1 applies with small epsilon.
    Standard supervised-learning assumption underlying the surrogate losses L1 and L2 in Eqs. (4)-(5).
  • domain assumption Output missingness in the semi-supervised extension is MCAR or MAR, not MNAR, so classification without bias correction is valid.
    Stated in the 'Extension to Missing Labels Scenario' section and adopted from standard SSL literature (Chapelle et al. 2006).

how reviews work

0 comments
Cite this review

Pith. "Pith review of MIRRAMS: Learning Robust Tabular Models under Unseen Missingness Shifts." pith.science (2026). https://pith.science/paper/5PCBH2Y4

@misc{pith2026250708280,
  author       = {Pith},
  title        = {Pith review of: MIRRAMS: Learning Robust Tabular Models under Unseen Missingness Shifts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5PCBH2Y4}},
  note         = {Machine review of arXiv:2507.08280}
}
read the original abstract

The presence of missing values often reflects variations in data collection policies, which may shift across time or locations, even when the underlying feature distribution remains stable. Such shifts in the missingness distribution between training and test inputs pose a significant challenge to achieving robust predictive performance. In this study, we propose a novel deep learning framework designed to address this challenge, particularly in the common yet challenging scenario where the test-time dataset is unseen. We begin by introducing a set of mutual information-based conditions, called MI robustness conditions, which guide the prediction model to extract label-relevant information. This promotes robustness against distributional shifts in missingness at test-time. To enforce these conditions, we design simple yet effective loss terms that collectively define our final objective, called MIRRAMS. Importantly, our method does not rely on any specific missingness assumption such as MCAR, MAR, or MNAR, making it applicable to a broad range of scenarios. Furthermore, it can naturally extend to cases where labels are also missing in training data, by generalizing the framework to a semi-supervised learning setting. Extensive experiments across multiple benchmark tabular datasets demonstrate that MIRRAMS consistently outperforms existing state-of-the-art baselines and maintains stable performance under diverse missingness conditions. Moreover, it achieves superior performance even in fully observed settings, highlighting MIRRAMS as a powerful, off-the-shelf framework for general-purpose tabular learning.

Figures

Figures reproduced from arXiv: 2507.08280 by the authors.

Figure 1
Figure 1. An illustration of the implementation of MIR [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of averaged test AUC scores (%) across 10 tabular datasets under settings with MCAR output missingness and MAR [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. (1st to 4th) Test AUC scores for varying values of r, λ1, λ2, and τ , respectively. (5th) Running time comparison. Each runtime is rescaled relative to that of MIRRAMS. additional masking during training acts as an implicit regu￾larizer, preventing the model from extracting excessive in￾formation that may lead to overfitting. This conjecture is empirically supported: setting either λ1 or λ2 to zero re￾sults in lower… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Data Language Models: A New Foundation Model Class for Tabular Data

    cs.AI 2026-05 unverdicted novelty 7.0 of 10

    Schema-1 is the first Data Language Model that natively understands raw tabular data and outperforms gradient-boosted ensembles, AutoML, and prior tabular foundation models on row-level prediction and imputation tasks.

  2. LUCAS-MEGA: A Large-Scale Multimodal Dataset for Representation Learning in Soil-Environment Systems

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    LUCAS-MEGA fuses 68 soil-environment datasets into a 70k-sample multimodal resource that supports self-supervised pretraining of SoilFormer, whose representations align with known soil processes.

  3. LUCAS-MEGA: A Large-Scale Multimodal Dataset for Representation Learning in Soil-Environment Systems

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    LUCAS-MEGA fuses 68 heterogeneous soil datasets into a 70k-sample multimodal collection and demonstrates its value by pretraining a tabular transformer whose representations align with established soil processes.

Reference graph

Works this paper leans on

39 extracted references · 27 canonical work pages · cited by 2 Pith papers

  1. [1]

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun - Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020 , 2020

    work page 2020

  2. [2]

    Ehrtemporalvariability: delineating temporal data-set shifts in electronic health records

    Carlos S \'a ez-Alba Guti \'e rrez-Sacrist \'a n-Isaac and Kohane-Juan Garc \' a-G \'o mez-Paul Avillach. Ehrtemporalvariability: delineating temporal data-set shifts in electronic health records. GigaScience--2020

    work page 2020

  3. [3]

    Domain adaptation under missingness shift

    Helen Zhou, Sivaraman Balakrishnan, and Zachary Lipton. Domain adaptation under missingness shift. In International Conference on Artificial Intelligence and Statistics , pages 9577--9606. PMLR, 2023

    work page 2023

  4. [4]

    Domain Adaptation Under MNAR Missingness

    Tyrel Stokes, Hyungrok Do, Saul Blecker, Rumi Chunara, and Samrachana Adhikari. Domain adaptation under mnar missingness. arXiv preprint arXiv:2504.00322 , 2025

    work page Pith review arXiv 2025

  5. [5]

    Managing dataset shift by adversarial validation for credit scoring

    Hongyi Qian, Baohui Wang, Ping Ma, Lei Peng, Songfeng Gao, and You Song. Managing dataset shift by adversarial validation for credit scoring. In Pacific Rim International Conference on Artificial Intelligence , pages 477--488. Springer, 2022

    work page 2022

  6. [6]

    Response rates, nonresponse bias, and data quality: results from a national survey of senior healthcare leaders

    Mark Meterko, Joseph D Restuccia, Kelly Stolzmann, David Mohr, Caitlin Brennan, Justin Glasgow, and Peter Kaboli. Response rates, nonresponse bias, and data quality: results from a national survey of senior healthcare leaders. Public Opinion Quarterly , 79(1):130--144, 2015

    work page 2015

  7. [7]

    Benchmarking distribution shift in tabular data with tableshift

    Josh Gardner, Zoran Popovic, and Ludwig Schmidt. Benchmarking distribution shift in tabular data with tableshift. Advances in Neural Information Processing Systems , 36:53385--53432, 2023

    work page 2023

  8. [8]

    A review and suggestions for synthetic data generation strategies using deep generative models

    Jiwoo Kim, Sunghoon Kwon, and Dongha Kim. A review and suggestions for synthetic data generation strategies using deep generative models. The Korean Data & Information Science Society , 34(5):791--810, 2023

    work page 2023

Show all 39 references
  1. [9]

    Embeddings for tabular data: A survey

    Rajat Singh and Srikanta Bedathur. Embeddings for tabular data: A survey. arXiv preprint arXiv:2302.11777 , 2023

    2023 arXiv

  2. [10]

    Scarf: Self-supervised contrastive learning using random feature corruption

    Dara Bahri, Heinrich Jiang, Yi Tay, and Donald Metzler. Scarf: Self-supervised contrastive learning using random feature corruption. arXiv preprint arXiv:2106.15147 , 2021

    2021 arXiv

  3. [11]

    Recontab: Regularized contrastive representation learning for tabular data

    Suiyao Chen, Jing Wu, Naira Hovakimyan, and Handong Yao. Recontab: Regularized contrastive representation learning for tabular data. arXiv preprint arXiv:2310.18541 , 2023

    2023 arXiv

  4. [12]

    Switchtab: Switched autoencoders are effective tabular learners

    Jing Wu, Suiyao Chen, Qi Zhao, Renat Sergazinov, Chen Li, Shengjie Liu, Chongchao Zhao, Tianpei Xie, Hanqing Guo, Cheng Ji, et al. Switchtab: Switched autoencoders are effective tabular learners. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pag...

    2024

  5. [13]

    Revisiting deep learning models for tabular data

    Yury Gorishniy, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Revisiting deep learning models for tabular data. In Advances in Neural Information Processing Systems 34: , pages 18932--18943, 2021

    2021

  6. [14]

    Xtab: Cross-table pretraining for tabular transformers

    Bingzhao Zhu, Xingjian Shi, Nick Erickson, Mu Li, George Karypis, and Mahsa Shoaran. Xtab: Cross-table pretraining for tabular transformers. arXiv preprint arXiv:2305.06090 , 2023

    2023 arXiv

  7. [15]

    Tabpfn: A transformer that solves small tabular classification problems in a second

    Noah Hollmann, Samuel M \"u ller, Katharina Eggensperger, and Frank Hutter. Tabpfn: A transformer that solves small tabular classification problems in a second. arXiv preprint arXiv:2207.01848 , 2022

    2022 arXiv

  8. [16]

    Tabtransformer: Tabular data modeling using contextual embeddings

    Xin Huang, Ashish Khetan, Milan Cvitkovic, and Zohar Karnin. Tabtransformer: Tabular data modeling using contextual embeddings. arXiv preprint arXiv:2012.06678 , 2020

    2012 arXiv

  9. [17]

    Bayan Bruss, and Tom Goldstein

    Gowthami Somepalli, Micah Goldblum, Avi Schwarzschild, C. Bayan Bruss, and Tom Goldstein. SAINT: improved neural networks for tabular data via row attention and contrastive pre-training. CoRR , abs/2106.01342, 2021

    2021 arXiv

  10. [18]

    VIME: extending the success of self- and semi-supervised learning to tabular domain

    Jinsung Yoon, Yao Zhang, James Jordon, and Mihaela van der Schaar. VIME: extending the success of self- and semi-supervised learning to tabular domain. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, Neu...

    2020

  11. [19]

    Unitabe: A universal pretraining protocol for tabular foundation model in data science

    Yazheng Yang, Yuqi Wang, Guang Liu, Ledell Wu, and Qi Liu. Unitabe: A universal pretraining protocol for tabular foundation model in data science. arXiv preprint arXiv:2307.09249 , 2023

    2023 arXiv

  12. [20]

    To predict or not to predict? proportionally masked autoencoders for tabular data imputation

    Jungkyu Kim, Kibok Lee, and Taeyoung Park. To predict or not to predict? proportionally masked autoencoders for tabular data imputation. In Toby Walsh, Julie Shah, and Zico Kolter, editors, AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, F...

    2025

  13. [21]

    Girshick

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll \' a r, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022 , pages 15979--1598...

    2022

  14. [22]

    Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, D...

    2020

  15. [23]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 , 2018

    2018 arXiv

  16. [24]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 , 2014

    2014 arXiv

  17. [25]

    Dash: Semi-supervised learning with dynamic thresholding

    Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yufeng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 20...

    2021

  18. [26]

    Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling

    Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In Marc'Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaug...

    2021

  19. [27]

    Class-imbalanced semi-supervised learning with adaptive thresholding

    Lan - Zhe Guo and Yufeng Li. Class-imbalanced semi-supervised learning with adaptive thresholding. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesv \' a ri, Gang Niu, and Sivan Sabato, editors, International Conference on Machine Learning, ICML 2022, 17-23 July ...

    2022

  20. [28]

    Freematch: Self-adaptive thresholding for semi-supervised learning

    Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, Bernt Schiele, and Xing Xie. Freematch: Self-adaptive thresholding for semi-supervised learning. In The Eleventh International Conference on Learnin...

    2023

  21. [29]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

    2017

  22. [30]

    Random forests

    Leo Breiman. Random forests. Machine learning , 45:5--32, 2001

    2001

  23. [31]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining , pages 785--794, 2016

    2016

  24. [32]

    Catboost: unbiased boosting with categorical features

    Liudmila Ostroumova Prokhorenkova, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, and Andrey Gulin. Catboost: unbiased boosting with categorical features. In Samy Bengio, Hanna M. Wallach, Hugo Larochelle, Kristen Grauman, Nicol \` o Cesa - Bianchi, and Roman Garnett, ...

    2018

  25. [33]

    o lkopf. Correcting sample selection bias by unlabeled data. In Bernhard Sch \

    Jiayuan Huang, Alexander J. Smola, Arthur Gretton, Karsten M. Borgwardt, and Bernhard Sch \" o lkopf. Correcting sample selection bias by unlabeled data. In Bernhard Sch \" o lkopf, John C. Platt, and Thomas Hofmann, editors, Advances in Neural Information Processing Systems 1...

    2006

  26. [34]

    Covariate shift by kernel mean matching

    Arthur Gretton, Alex Smola, Jiayuan Huang, Marcel Schmittfull, Karsten Borgwardt, Bernhard Sch \"o lkopf, et al. Covariate shift by kernel mean matching. Dataset shift in machine learning , 3(4):5, 2009

    2009

  27. [35]

    Analysis of kernel mean matching under covariate shift

    Yaoliang Yu and Csaba Szepesv \' a ri. Analysis of kernel mean matching under covariate shift. In Proceedings of the 29th International Conference on Machine Learning, ICML 2012, Edinburgh, Scotland, UK, June 26 - July 1, 2012 . icml.cc / Omnipress, 2012

    2012

  28. [36]

    Lempitsky

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, Fran c ois Laviolette, Mario Marchand, and Victor S. Lempitsky. Domain-adversarial training of neural networks. CoRR , abs/1505.07818, 2015

    2015 arXiv

  29. [37]

    Adversarial discriminative domain adaptation

    Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 , pages 2962--2971. IEEE Computer Society, 2017

    2017

  30. [38]

    Learning semantic representations for unsupervised domain adaptation

    Shaoan Xie, Zibin Zheng, Liang Chen, and Chuan Chen. Learning semantic representations for unsupervised domain adaptation. In Jennifer G. Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsm \" a ssan, Sto...

    2018

  31. [39]

    An intuitive proof of the data processing inequality

    Normand J Beaudry and Renato Renner. An intuitive proof of the data processing inequality. arXiv preprint arXiv:1107.0740 , 2011

    2011 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.