Pith. sign in

REVIEW 4 major objections 4 minor 56 references

Improving realistic semi-supervised learning with doubly robust estimation

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read Explicitly estimating the unlabeled class distribution with a doubly robust estimator and plugging it into existing pseudo-labeling methods improves accuracy in realistic long-tailed semi-supervised learning.

desk verdict The two-stage plug-in idea is sensible and the empirical gains look real, but the doubly robust theory does not survive contact with label shift. read the letter →

arxiv 2502.00279 v1 pith:MXMRBQ2Y submitted 2025-02-01 cs.LG stat.ML

classification cs.LGstat.ML MSC 62F1268T0562D10
keywords semi-supervisedlearninglong-taileddistributionlabelshiftdoublyrobustestimationpseudo-labelingmissingnessmechanismexpectation-maximizationclass
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

In realistic long-tailed semi-supervised learning, the class distribution of the unlabeled data is unknown and differs from the labeled data, and pseudo-labeling methods that estimate it on the fly from their own predictions are biased toward head classes. This paper claims that the unlabeled class distribution can be estimated explicitly, as a standalone first step, with a doubly robust estimator that is asymptotically efficient under standard nuisance-convergence assumptions. The estimate is then frozen and plugged into existing pseudo-labeling methods (SimPro and BOAT), which the experiments show improves accuracy in 9 of 10 CIFAR-10 settings, as well as on STL-10 and ImageNet-127. The paper also shows that SimPro is a reparameterization of a semi-supervised EM algorithm with label shift, unifying the two views. If the claim holds, SSL methods that assume known or label-matched unlabeled distributions can be turned into robust long-tailed learners simply by pre-estimating the shift.

What carries the argument

The load-bearing object is the doubly robust estimator of the finite-dimensional parameter $P(Y)$, written as a dataset average of the outcome-regression term plus an inverse-probability-weighted residual, $\frac{1}{N}\sum_i [\hat{P}(Y=c|X_i) + \frac{\mathbf{1}(A_i=1)}{\hat{P}(A_i=1|Y_i)}(\mathbf{1}(Y_i=c) - \hat{P}(Y=c|X_i))]$. Its double robustness makes the first-stage estimate resilient to misspecification of either the classifier or the missingness mechanism, and the efficiency theorem turns the two nuisance errors into a second-order product so that the bias vanishes at an $n^{-1/2}$ rate. The second-stage plug-in is the parameter substitution into logit adjustment and pseudo-labeling within SimPro and BOAT.

What would settle it

Run the two-stage procedure on a controlled dataset where the classifier is deliberately misspecified or the labeled set is so small that the $n^{-1/4}$ rate fails; if the doubly robust estimate's total variation distance to the true unlabeled distribution stops beating the on-the-fly estimates, the practical bite of the efficiency claim is falsified. Alternatively, use cross-fitting and check whether the TV distance and downstream accuracy improve as Theorem 3.2's sample-splitting requirement would predict.

Watch

Extended reading notes

Core claim

The central claim is that explicitly estimating the unlabeled class distribution $P(Y|A=0)$ with the doubly robust estimator $\Psi_{\mathrm{dr}}$ yields a better estimate than on-the-fly pseudo-label-based estimates, and that freezing this estimate and plugging it into existing methods improves final classification accuracy. The estimator combines an outcome-regression term (the average model prediction) with an inverse-probability-weighting correction term that up-weights labeled data by the missingness mechanism; it is unbiased if either nuisance function is correct, and Theorem 3.2 asserts it is asymptotically normal with the efficient influence function variance when both converge at an $n^{-1/4}$ rate. Empirically, the plug-in versions SimPro+ and BOAT+ beat their bases in most settings, and the benefit is most pronounced in the consistent and uniform shift settings where SimPro alone overshoots head classes.

Load-bearing premise

The strong guarantee rests on the assumption that both the classifier and the label-selection mechanism converge at an $n^{-1/4}$ rate and are estimated on separate data, but the experiments train them on the same data without splitting.

Editorial extensions

If this is right

  • SimPro+ improves accuracy over SimPro in 9 of 10 CIFAR-10 settings, and BOAT+ improves over BOAT in 8 of 10 settings.
  • The plug-in gains extend to STL-10 and ImageNet-127, and a small Wide ResNet-28-2 suffices for stage-1 distribution estimation.
  • Even when stage-1 estimation is mediocre on CIFAR-100, the plug-in does not degrade accuracy and reaches parity with the baselines in most settings.
  • Under the stated assumptions, the unlabeled class distribution estimate is asymptotically efficient, meaning no regular estimator can beat its limiting variance.
  • Because only a finite-dimensional parameter is estimated in stage 1, the preliminary step is computationally cheap and decouples from the final classifier size.

Reading between the lines

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

  • A testable extension is to apply the same two-stage plug-in to any pseudo-labeling SSL method that accepts a class-prior argument, such as DASO or CReST+, predicting that their head-class overshoot will drop whenever the DR estimate is better than the online estimate.
  • The theory assumes sample splitting, but the experiments train the nuisance functions on the same data without cross-fitting; whether the practical gains persist and whether efficiency holds without the split is an open question the paper leaves implicit.
  • Because the estimator targets $P(Y)$ while the test distribution is assumed uniform, the approach could be extended to jointly estimate $P(Y|A=0)$ and adapt to test-time label shift, connecting to calibration-style post-hoc corrections.
  • An implication of the efficiency result is that stage-1 could be swapped for any nuisance estimator meeting the $n^{-1/4}$ rate, making the method a general wrapper around label-shift SSL rather than a change specific to SimPro or BOAT.
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

4 major / 4 minor

Summary. The paper tackles realistic long-tailed semi-supervised learning (RTSSL), where the unlabeled class distribution P(Y|A=0) is unknown and differs from the labeled distribution. The authors propose a two-stage procedure: stage 1 estimates P(Y|A=0) via an EM algorithm and a 'doubly robust' estimator Ψdr that combines a classifier P(Y|X,θ) and a missingness mechanism P(A|Y,θ); stage 2 plugs this frozen estimate into existing pseudo-labeling methods (SimPro, BOAT). The central theoretical claim, Theorem 3.2 in Section 3.3.1, asserts that Ψdr is asymptotically normal and efficient under n^{-1/4} nuisance convergence. Experiments on CIFAR-10, CIFAR-100, STL-10, and ImageNet-127 report improved class-distribution estimates and improved classification accuracy in most settings.

Significance. The empirical framework is appealing: the connection between RTSSL and non-ignorable missing data is natural, the two-stage plug-in idea is simple and reproducible, and the experimental comparison is broad, including multiple baselines, ablations, and several label-shift configurations. If the theoretical guarantee were valid, the paper would make a useful contribution to semiparametric SSL. However, the central theoretical result is not established, and the estimator does not actually possess the advertised double-robustness property. Since the abstract and Section 3.3 present this theory as a primary contribution, the flaw substantially reduces the significance of the work; the empirical gains alone are modest and would require reframing.

major comments (4)
  1. [Appendix A, Eq. (20)] The proof of Theorem 3.2 relies on an invalid algebraic step. Under the paper's label-shift assumption A ⟂ X | Y (Eq. 2), the term E[1(A=1)/Pθ(A=1|Y) 1(Y=c)] equals E[P(A=1|Y)/Pθ(A=1|Y) 1(Y=c)] = E[r(Y)/rθ(Y) 1(Y=c)] = r(c)/rθ(c) P(Y=c). The paper instead replaces this with E[P(A=1|Y)/Pθ(A=1|Y) P(Y=c|X)], treating the propensity ratio as if it were a function of X alone. That replacement would require missing-at-random structure (A ⟂ Y | X), not the label-shift assumption. Consequently the product-of-errors bound in Eq. (21) and the claimed o_P(N^{-1/2}) remainder do not follow, and the asymptotic normality and efficiency conclusion of Theorem 3.2 is unsupported.
  2. [Section 3.3, after Eq. (13)] The estimator Ψdr is not doubly robust under label shift. The text claims that if either P(Y|X) or P(A=1|Y) is correct, Ψdr is unbiased. The second direction fails: if Pθ(Y|X,θ)=P(Y|X) is correct but Pθ(A|Y,θ) is misspecified, then E[Ψdr(c)] = P(Y=c) + E[(P(A=1|Y)/Pθ(A=1|Y))(1(Y=c) - P(Y=c|X))], which is not zero in general. For example, with uninformative X, P(Y=1)=0.5, P(A=1|Y=1)=0.8, and Pθ(A=1|Y=1)=0.2, one obtains E[Ψdr(1)] ≈ 1.59 ≠ 0.5. Only the 'propensity correct' direction is unbiased. This directly contradicts the abstract's claim of a 'strong theoretical guarantee' and the double-robustness statement in Section 3.3.
  3. [Section 4.1 / Appendix C / Assumption A.1.3] The experiments do not implement the sample splitting required by Assumption A.1.3. The stage-1 Wide ResNet-28-2 nuisances are trained by EM on the same combined dataset used for the sample average in Eq. (13), without cross-fitting or auxiliary samples, and no evidence is provided that the EM-trained estimates achieve the n^{-1/4} L2 rates of Assumption 3.1. Thus, even if Theorem 3.2 were correctly proven, the experiments would not instantiate its conditions, and the claim that a small network suffices for stage 1 is not supported by the stated theory.
  4. [Appendix A, Eq. (27)] The efficiency argument also contains an error in the tangent-space decomposition. The paper defines T_{A|Y} = {h(A,Y) : E[h|Y]=0} and states that the second term in Eq. (27), namely (1(A=1)/P(A=1|Y)-1)(1(Y=c)-P(Y=c|X)), lies in T_{A|Y}. This term depends on X through P(Y=c|X), so it is not a function of (A,Y) alone and cannot belong to T_{A|Y} as defined. The projection of the X-dependent factor onto functions of Y would be needed, but this is not done. Hence the conclusion that φ is the efficient influence function is not established.
minor comments (4)
  1. [Section 3.2, Eq. (9)] The logit-adjustment loss in Eq. (9) is written with a '+' between P(Y=c|X,xi,uniform,θ) and P(Y=c), which is not the standard logit adjustment of Menon et al.; the formula appears to be a typesetting error and should be corrected to the usual multiplicative or logit-space form.
  2. [Section 4.2, text after Table 4] The text says the approach 'does not degrade the accuracy for most settings' on CIFAR-100, but Table 4 shows that SimPro+ is worse than SimPro in 5 of 10 settings and BOAT+ is worse than BOAT in 4 of 10 settings (with 2 ties). The wording should be adjusted to match the table counts.
  3. [Section 3.3, Eq. (10)] The outcome-regression estimator Ψor averages P(Y=c|X,θ) over both labeled and unlabeled samples, which estimates the combined class distribution P(Y), not directly P(Y|A=0). The text should clarify this intermediate step and how P(Y|A=0) is recovered from the known P(A) and P(Y|A=1).
  4. [Throughout] There are several typographical and grammatical issues, such as 'as an initial step ,' in the abstract, 'we propose ... to estimate ... as an initial step, using ...' with an awkward comma, and 'shows that we can plug in this estimation' in Section 2. A careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DR estimator is a standard semiparametric plug-in evaluated against independent baselines; self-citations are not load-bearing.

full rationale

The paper's derivation chain is self-contained. The target P(Y|A=0) is estimated by a doubly robust combination (Eq. 13) of two nuisance functions P(Y|X,theta) and P(A|Y,theta) that are themselves learned by EM; using the same model's outputs as nuisances is standard semiparametric practice, and the estimator is not defined in terms of the target quantity. The claimed empirical improvement is tested by plugging the frozen estimate into SimPro and BOAT and comparing against the original algorithms on held-out test accuracy (Tables 3-6), so the prediction is not fitted to the outcome it is said to predict. The only self-citations (Pham et al., 2023) appear as contextual references for DR estimation and training instability, not as evidence for the paper's main claims, so they do not form a load-bearing self-citation chain. The identifiability of the label shift model is cited to external work (Miller & Futoma, 2023; Sportisse et al., 2023), and the n^{-1/4} rate assumption is cited to external results (Chernozhukov et al., 2022). A possible issue in the proof of Theorem 3.2 (the conditional expectation step in Eq. 20 assumes a missing-at-random structure rather than the paper's label shift assumption) is a mathematical correctness concern, not a circularity concern, because it does not make any output equal to an input by construction.

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

The method introduces no new free parameters or hand-chosen constants beyond the hyperparameters inherited from SimPro; it follows the SimPro recipe for both stages. The load-bearing assumptions are the label shift model, the uniform test distribution, and the rate and sample-splitting conditions needed for the theoretical guarantee. No invented entities are added.

assumptions (4)
  • domain assumption Label shift assumption: P(X|Y,A) = P(X|Y) (Equation 2).
    Identifies the missingness mechanism and posterior weights; if unlabeled feature distributions shift in other ways, the estimator targets the wrong quantity.
  • domain assumption Test class distribution is uniform, P(Y|uniform) = 1/C.
    Used for logit adjustment and evaluation; the method is not designed for a different test distribution.
  • ad hoc to paper Assumption 3.1: both nuisance estimators converge at op(N^{-1/4}).
    Required for Theorem 3.2; cited from Chernozhukov et al. but not verified for the stage-1 network or training procedure.
  • ad hoc to paper Assumption A.1.3: nuisances estimated on auxiliary samples independent of the averaging sample.
    Used in the proof of Theorem 3.2; not implemented in the experiments, so the theory as stated does not cover the tested procedure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving realistic semi-supervised learning with doubly robust estimation." pith.science (2026). https://pith.science/paper/MXMRBQ2Y

@misc{pith2026250200279,
  author       = {Pith},
  title        = {Pith review of: Improving realistic semi-supervised learning with doubly robust estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MXMRBQ2Y}},
  note         = {Machine review of arXiv:2502.00279}
}
read the original abstract

A major challenge in Semi-Supervised Learning (SSL) is the limited information available about the class distribution in the unlabeled data. In many real-world applications this arises from the prevalence of long-tailed distributions, where the standard pseudo-label approach to SSL is biased towards the labeled class distribution and thus performs poorly on unlabeled data. Existing methods typically assume that the unlabeled class distribution is either known a priori, which is unrealistic in most situations, or estimate it on-the-fly using the pseudo-labels themselves. We propose to explicitly estimate the unlabeled class distribution, which is a finite-dimensional parameter, \emph{as an initial step}, using a doubly robust estimator with a strong theoretical guarantee; this estimate can then be integrated into existing methods to pseudo-label the unlabeled data during training more accurately. Experimental results demonstrate that incorporating our techniques into common pseudo-labeling approaches improves their performance.

Figures

Figures reproduced from arXiv: 2502.00279 by the authors.

Figure 1
Figure 1. The labeled class distribution and 5 possible unlabeled class distributions studied in (Du et al., 2024). SimPro significantly overestimates the head classes in consistent, reverse and head-tail settings. Our doubly-robust estimate is more accurate at the head classes as well as the overall distribution in all but the middle setting, as measured in total variation distance in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our 2-stage method (Section 3.3). In stage 1, we use Expectation-Maximization (EM, Section 3.1) to estimate the missingness mechanism and classifier from observable data. These quantities are used as nuisance components for the doubly-robust estimator of the class distribution Equation (13). In stage 2, we can use EM or other existing methods that also use logit-adjustment with the (unlabeled) class dist… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 36 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Maximum likelihood with bias-corrected calibration is hard-to-beat at label shift adaptation

    Alexandari, A., Kundaje, A., and Shrikumar, A. Maximum likelihood with bias-corrected calibration is hard-to-beat at label shift adaptation. In International Conference on Machine Learning, pp.\ 222--232. PMLR, 2020

  3. [3]

    E., and McGuinness, K

    Arazo, E., Ortego, D., Albert, P., O’Connor, N. E., and McGuinness, K. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In 2020 International joint conference on neural networks (IJCNN), pp.\ 1--8. IEEE, 2020

  4. [4]

    Regularized learning for domain adaptation under label shifts

    Azizzadenesheli, K., Liu, A., Yang, F., and Anandkumar, A. Regularized learning for domain adaptation under label shifts. arXiv preprint arXiv:1903.09734, 2019

  5. [5]

    D., Kurakin, A., Sohn, K., Zhang, H., and Raffel, C

    Berthelot, D., Carlini, N., Cubuk, E. D., Kurakin, A., Sohn, K., Zhang, H., and Raffel, C. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019 a

  6. [6]

    Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., and Raffel, C. A. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019 b

  7. [7]

    Buda, M., Maki, A., and Mazurowski, M. A. A systematic study of the class imbalance problem in convolutional neural networks. Neural networks, 106: 0 249--259, 2018

  8. [8]

    Semi-supervised learning (chapelle, o

    Chapelle, O., Scholkopf, B., and Zien, A. Semi-supervised learning (chapelle, o. et al., eds.; 2006)[book reviews]. IEEE Transactions on Neural Networks, 20 0 (3): 0 542--542, 2009

Show all 56 references
  1. [9]

    Double/debiased machine learning for treatment and structural parameters, 2018

    Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., and Robins, J. Double/debiased machine learning for treatment and structural parameters, 2018

  2. [10]

    M., and Syrgkanis, V

    Chernozhukov, V., Newey, W., Quintas-Mart nez, V. M., and Syrgkanis, V. Riesznet and forestriesz: Automatic debiased machine learning with neural nets and random forests. In International Conference on Machine Learning, pp.\ 3901--3914. PMLR, 2022

  3. [11]

    An analysis of single-layer networks in unsupervised feature learning

    Coates, A., Ng, A., and Lee, H. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pp.\ 215--223. JMLR Workshop and Conference Proceedings, 2011

  4. [12]

    Class-balanced loss based on effective number of samples

    Cui, Y., Jia, M., Lin, T.-Y., Song, Y., and Belongie, S. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 9268--9277, 2019

  5. [13]

    Simpro: A simple probabilistic framework towards realistic long-tailed semi-supervised learning

    Du, C., Han, Y., and Huang, G. Simpro: A simple probabilistic framework towards realistic long-tailed semi-supervised learning. arXiv preprint arXiv:2402.13505, 2024

  6. [14]

    Rda: Reciprocal distribution alignment for robust semi-supervised learning

    Duan, Y., Qi, L., Wang, L., Zhou, L., and Shi, Y. Rda: Reciprocal distribution alignment for robust semi-supervised learning. In European Conference on Computer Vision, pp.\ 533--549. Springer, 2022

  7. [15]

    Towards semi-supervised learning with non-random missing labels

    Duan, Y., Zhao, Z., Qi, L., Zhou, L., Wang, L., and Shi, Y. Towards semi-supervised learning with non-random missing labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 16121--16131, 2023

  8. [16]

    Cossl: Co-learning of representation and classifier for imbalanced semi-supervised learning

    Fan, Y., Dai, D., Kukleva, A., and Schiele, B. Cossl: Co-learning of representation and classifier for imbalanced semi-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 14574--14584, 2022

  9. [17]

    Foster, D. J. and Syrgkanis, V. Orthogonal statistical learning. The Annals of Statistics, 51 0 (3): 0 879--908, 2023

  10. [18]

    Boosting consistency in dual training for long-tailed semi-supervised learning

    Gan, K., Wei, T., and Zhang, M.-L. Boosting consistency in dual training for long-tailed semi-supervised learning. arXiv preprint arXiv:2406.13187, 2024

  11. [19]

    A unified view of label shift estimation

    Garg, S., Wu, Y., Balakrishnan, S., and Lipton, Z. A unified view of label shift estimation. Advances in Neural Information Processing Systems, 33: 0 3290--3300, 2020

  12. [20]

    and Bengio, Y

    Grandvalet, Y. and Bengio, Y. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004

  13. [21]

    On non-random missing labels in semi-supervised learning

    Hu, X., Niu, Y., Miao, C., Hua, X.-S., and Zhang, H. On non-random missing labels in semi-supervised learning. arXiv preprint arXiv:2206.14923, 2022

  14. [22]

    Ibrahim, J. G. and Lipsitz, S. R. Parameter estimation from incomplete data in binomial regression when the missing data mechanism is nonignorable. Biometrics, pp.\ 1071--1078, 1996

  15. [23]

    Deepmatch: Balancing deep covariate representations for causal inference using adversarial training

    Kallus, N. Deepmatch: Balancing deep covariate representations for causal inference using adversarial training. In International Conference on Machine Learning, pp.\ 5067--5077. PMLR, 2020

  16. [24]

    Kennedy, E. H. Towards optimal doubly robust estimation of heterogeneous causal effects. Electronic Journal of Statistics, 17 0 (2): 0 3008--3049, 2023

  17. [25]

    Kennedy, E. H. Semiparametric doubly robust targeted double machine learning: a review. Handbook of Statistical Methods for Precision Medicine, pp.\ 207--236, 2024

  18. [26]

    J., and Shin, J

    Kim, J., Hur, Y., Park, S., Yang, E., Hwang, S. J., and Shin, J. Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning. Advances in neural information processing systems, 33: 0 14567--14579, 2020

  19. [27]

    and Hinton, G

    Krizhevsky, A. and Hinton, G. Learning multiple layers of features from tiny images. Technical Report 0, University of Toronto, Toronto, Ontario, 2009. URL https://www.cs.toronto.edu/ kriz/learning-features-2009-TR.pdf

  20. [28]

    and Aila, T

    Laine, S. and Aila, T. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016

  21. [29]

    Lee, D.-H. et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, pp.\ 896. Atlanta, 2013

  22. [30]

    Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning

    Lee, H., Shin, S., and Kim, H. Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning. Advances in Neural Information Processing Systems, 34: 0 7082--7094, 2021

  23. [31]

    Detecting and correcting for label shift with black box predictors

    Lipton, Z., Wang, Y.-X., and Smola, A. Detecting and correcting for label shift with black box predictors. In International conference on machine learning, pp.\ 3122--3130. PMLR, 2018

  24. [32]

    Three heads are better than one: Complementary experts for long-tailed semi-supervised learning

    Ma, C., Elezi, I., Deng, J., Dong, W., and Xu, C. Three heads are better than one: Complementary experts for long-tailed semi-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 14229--14237, 2024

  25. [33]

    K., Jayasumana, S., Rawat, A

    Menon, A. K., Jayasumana, S., Rawat, A. S., Jain, H., Veit, A., and Kumar, S. Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314, 2020

  26. [34]

    Miller, A. C. and Futoma, J. Label shift estimators for non-ignorable missing data. arXiv preprint arXiv:2310.18261, 2023

  27. [35]

    Neal, R. M. and Hinton, G. E. A view of the em algorithm that justifies incremental, sparse, and other variants. In Learning in graphical models, pp.\ 355--368. Springer, 1998

  28. [36]

    Q., Kim, H., and Sabuncu, M

    Nguyen, M., Wang, A. Q., Kim, H., and Sabuncu, M. R. Adapting to shifting correlations with unlabeled data calibration. arXiv preprint arXiv:2409.05996, 2024

  29. [37]

    Oh, Y., Kim, D.-J., and Kweon, I. S. DASO : Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 9786--9796, 2022

  30. [38]

    A., Huynh-Pham, P.-M., Santacatterina, M., Lim, S.-N., and Zabih, R

    Pham, K., Hirshberg, D. A., Huynh-Pham, P.-M., Santacatterina, M., Lim, S.-N., and Zabih, R. Stable estimation of survival causal effects. arXiv preprint arXiv:2310.02278, 2023

  31. [39]

    Balanced meta-softmax for long-tailed visual recognition

    Ren, J., Yu, C., Ma, X., Zhao, H., Yi, S., et al. Balanced meta-softmax for long-tailed visual recognition. Advances in neural information processing systems, 33: 0 4175--4186, 2020

  32. [40]

    Rubin, D. B. Inference and missing data. Biometrika, 63 0 (3): 0 581--592, 1976

  33. [41]

    Adjusting the outputs of a classifier to new a priori probabilities: a simple procedure

    Saerens, M., Latinne, P., and Decaestecker, C. Adjusting the outputs of a classifier to new a priori probabilities: a simple procedure. Neural computation, 14 0 (1): 0 21--41, 2002

  34. [42]

    Don't fear the unlabelled: safe semi-supervised learning via simple debiasing

    Schmutz, H., Humbert, O., and Mattei, P.-A. Don't fear the unlabelled: safe semi-supervised learning via simple debiasing. arXiv preprint arXiv:2203.07512, 2022

  35. [43]

    Adapting neural networks for the estimation of treatment effects

    Shi, C., Blei, D., and Veitch, V. Adapting neural networks for the estimation of treatment effects. Advances in neural information processing systems, 32, 2019

  36. [44]

    A., Cubuk, E

    Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C. A., Cubuk, E. D., Kurakin, A., and Li, C.-L. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33: 0 596--608, 2020

  37. [45]

    Are labels informative in semi-supervised learning? estimating and leveraging the missing-data mechanism

    Sportisse, A., Schmutz, H., Humbert, O., Bouveyron, C., and Mattei, P.-A. Are labels informative in semi-supervised learning? estimating and leveraging the missing-data mechanism. In International Conference on Machine Learning, pp.\ 32521--32539. PMLR, 2023

  38. [46]

    P., Ebrahimi, S., and D'Amour, A

    Sun, Q., Murphy, K. P., Ebrahimi, S., and D'Amour, A. Beyond invariance: test-time label-shift adaptation for addressing ``spurious''' correlations. Advances in Neural Information Processing Systems, 36: 0 23789--23812, 2023

  39. [47]

    Tsiatis, A. A. Semiparametric theory and missing data, volume 4. Springer, 2006

  40. [48]

    Van der Vaart, A. W. Asymptotic statistics, volume 3. Cambridge university press, 2000

  41. [49]

    and Athey, S

    Wager, S. and Athey, S. Estimation and inference of heterogeneous treatment effects using random forests. Journal of the American Statistical Association, 113 0 (523): 0 1228--1242, 2018

  42. [50]

    Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning

    Wei, C., Sohn, K., Mellina, C., Yuille, A., and Yang, F. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10857--10866, 2021

  43. [51]

    and Gan, K

    Wei, T. and Gan, K. Towards realistic long-tailed semi-supervised learning: Consistency is all you need. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 3469--3478, 2023

  44. [52]

    Learning label shift correction for test-agnostic long-tailed recognition

    Wei, T., Mao, Z., Zhou, Z.-H., Wan, Y., and Zhang, M.-L. Learning label shift correction for test-agnostic long-tailed recognition. In Forty-first International Conference on Machine Learning, 2024

  45. [53]

    Xie, Q., Luong, M.-T., Hovy, E., and Le, Q. V. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10687--10698, 2020

  46. [54]

    Towards causal foundation model: on duality between causal inference and attention

    Zhang, J., Jennings, J., Hilmkil, A., Pawlowski, N., Zhang, C., and Ma, C. Towards causal foundation model: on duality between causal inference and attention. arXiv preprint arXiv:2310.00809, 2023

  47. [55]

    Dc-ssl: Addressing mismatched class distribution in semi-supervised learning

    Zhao, Z., Zhou, L., Duan, Y., Wang, L., Qi, L., and Shi, Y. Dc-ssl: Addressing mismatched class distribution in semi-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 9757--9765, 2022

  48. [56]

    Doubly-robust self-training

    Zhu, B., Ding, M., Jacobson, P., Wu, M., Zhan, W., Jordan, M., and Jiao, J. Doubly-robust self-training. Advances in Neural Information Processing Systems, 36, 2024

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.