Pith. sign in

REVIEW 3 major objections 5 minor 72 references

The Label Complexity of Class-Conditional Coverage under Distribution Shift

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Per-class coverage under joint distribution shift is unidentifiable for label-free methods; the minimax cost of restoring it with classwise thresholds is Θ(ε⁻² log K) target labels per class.

desk verdict A genuinely new label-complexity result with matching bounds, but the central proofs live in a missing appendix; referee should demand the supplementary file before judging. read the letter →

arxiv 2607.18088 v2 pith:QAHZQCXH submitted 2026-07-20 cs.LG cs.CV

Weijia Han , Lisha Qu This is my paper
classification cs.LGcs.CV
keywords labelcomplexityclass-conditionalcoveragedistributionshiftconformalpredictionper-classMondrianminimaxratePACvalidity
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 establishes that restoring per-class coverage under joint covariate-label distribution shift has an exact price in target labels, and that a marginal coverage certificate cannot substitute for it. On a real cross-subject skeleton action recognition benchmark, split conformal prediction keeps marginal coverage near ninety percent while the worst class is covered only about seventy percent of the time; the paper shows this is not a fixable algorithmic gap but an identifiability barrier. Because the target class-conditional score law is unidentified from labeled source data and an unlabeled target sample, no label-free method can be simultaneously valid and efficient per class uniformly over consistent target laws. For classwise threshold procedures, the per-class labels needed to recover every class quantile to tolerance ε with simultaneous PAC validity grow as Θ(ε⁻² log K), with matching upper and lower bounds. The paper also shows that pseudo-label estimators gain almost nothing where coverage collapses, while a label-free per-class calibration on the source recovers a substantial share of the gap whenever marginal coverage still holds.

What carries the argument

The central object is the target per-class quantile q*_c of the class-conditional score law F^T_c, along with the classwise threshold prediction rule that includes every class whose score falls below q_c. The impossibility result works by constructing indistinguishable target laws whose quantiles differ, forcing any universally valid label-free rule to be conservative. The positive result is carried by PAC Audit Mondrian, a classwise threshold procedure that inflates each per-class quantile by a finite-sample concentration slack and achieves simultaneous PAC validity and ε-accurate recovery with m_c = O(ε⁻² log(K/δ)) labels per class; a two-region construction shows the ε⁻² rate and the log

What would settle it

Construct a shift instance with fixed source joint distribution and fixed target covariate marginal, and two target label laws whose per-class quantiles differ by more than ε, and exhibit a label-free procedure that is per-class valid with expected set size within a small constant of the oracle under both laws — that would refute Proposition 2. Alternatively, on real data, deliberately make specific classes systematically harder at deployment while keeping the marginal shift mild, and show that source per-class calibration no longer lifts worst-class coverage, confirming the within-class excha

Watch

Extended reading notes

Core claim

The central claim is that per-class coverage under joint shift is a label-bounded problem with a sharp minimax rate, not a deficiency that clever labeling can remove. Proposition 2 constructs two target joint laws that share the same source joint distribution and the same target covariate marginal — hence are indistinguishable from all label-free observations — but whose per-class score quantiles differ by an explicit constant; consequently any label-free rule that is valid on every consistent target law must inflate its sets, and efficiency requires target labels. Theorem 1 proves that for classwise threshold predictors, the per-class audit labels needed for simultaneous (1−δ)-PAC validity

Load-bearing premise

The load-bearing premise is within-class exchangeability between source calibration scores and target test scores for the same class; if the shift also changes score distributions within a class, the label-free Mondrian recovery observed on real data would not follow, a limitation the paper acknowledges.

Editorial extensions

If this is right

  • A marginal coverage certificate alone is not evidence of per-class safety under distribution shift; a single global number can hide severe worst-class undercoverage.
  • Restoring per-class validity under joint shift requires target labels at a precisely characterized rate: Θ(ε⁻² log K) labels per class for simultaneous PAC-valid ε-accurate quantile recovery, and no label-free procedure can achieve this uniformly.
  • Unlabeled target data and pseudo-label estimators do not shortcut the label cost; prediction-powered estimators gain at most a small constant factor on the classes where coverage collapses.
  • Label-free per-class calibration on the source (Mondrian conformal prediction) recovers a substantial share of the per-class gap when within-class exchangeability approximately holds and marginal coverage is preserved, but stops being effective once marginal coverage itself breaks.

Reading between the lines

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

  • The matching rate Θ(ε⁻² log K) suggests a budget-constrained deployment could allocate target labels in proportion to per-class severity rather than uniformly; the paper does not derive such an allocation rule, so this is an extension.
  • The impossibility construction is generic — it only needs a score that separates two positive-probability regions — so a similar no-free-lunch result likely applies to other nested conditional guarantees, such as subgroup coverage by sensitive attributes.
  • The paper's severity spectrum hints that label-free recovery succeeds exactly when marginal coverage holds; a prospective test for predicting when this recovery will fail, built from label-free signals, is left for future work and could turn the after-the-fact pattern into a usable decision rule.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper studies per-class (class-conditional) validity of conformal prediction under distribution shift that acts jointly on covariates and labels. It proves (Prop. 2) a label-free impossibility: no method using only labeled source data and unlabeled target covariates can be simultaneously valid and efficient for every consistent target law. It gives a PAC audit scheme and shows (Thm. 1) that the per-class labeled target audit needed for simultaneous validity and ε-accurate quantile recovery is Θ(ε^{-2} log K) for classwise threshold procedures. A case study on NTU-60/N-UCLA and corrupted CIFAR shows marginal coverage can hold while per-class coverage collapses, and a source-only Mondrian calibration partially recovers the gap; a PPI-based pseudo-label ceiling is measured on two shifts.

Significance. If the deferred proofs are correct, the impossibility result and the matching label-complexity bounds are useful and substantially sharpen the folklore that 'marginal coverage is not enough' by giving an exact query-complexity characterization. The empirical part is valuable as a concrete demonstration, and the paper is unusually honest about its limitations (within-class exchangeability, no prospective decision rule, limited pseudo-label control span). The theory does not appear to fit free parameters to the target result. The main weakness is that all load-bearing constructions and proofs are deferred to an unavailable supplementary file.

major comments (3)
  1. [Section 4, Prop. 2 and Thm. 1(c)-(d)] The central lower-bound results cannot be checked from the submitted text. The two-point construction behind Prop. 2, the hard two-law instance in Thm. 1(c), and the K-region coupon-collector instance in Thm. 1(d) are all stated as 'in the supplementary material', and the checklist says full proofs are in the supplementary file; that file is not part of the manuscript under review. Since these are the load-bearing claims of the paper, please either append the complete proofs and constructions to the main submission or provide them as a supplement that is actually available to the referee, and state explicitly the constants δ, ε0, c0 and Cα used in the lower bounds.
  2. [Section 4, Thm. 1(d)] The statement 'Give a procedure n_c i.i.d. labeled target pairs from region c, with fixed quotas' is not a precise minimax formulation. Is the vector (n_c) chosen by the procedure before seeing data, is it a fixed budget chosen by the analyst, and does the lower bound hold for every such choice? What exactly is 'the instance' over which 'probability at least 3/4 under every law' is required? Please state Theorem 1(d) in standard minimax form: a class of procedures, a family of target laws, and the probability space (algorithm randomness, labeled and unlabeled draws). Without this, the lower-bound statement is not checkable.
  3. [Section 5, Remark 1] The practical recommendation that label-free source Mondrian 'recovers much of the gap' is conditional on within-class exchangeability of source and target scores, which the paper admits is violated under deployment shift. The within/between-class decomposition is left to future work. This is not a flaw in the theoretical results, but it means the case study does not establish when the empirical recovery will transfer. Please make this conditionality more prominent in the abstract/conclusion, where the recovery claim appears without the qualifier.
minor comments (5)
  1. [Sec. 4, Prop. 2 display] The displayed definition of β is garbled ('β= PT (A)(ε−α) ε>0,'); please restructure the statement so the two-point construction, the event A, and the level ε are defined before β.
  2. [Sec. 4, Thm. 1(a)] Please clarify the relationship between the ordinary conformal non-vacuity threshold m_c≥ceil(1/α)-1 (m=9 at α=0.1) and PAC Audit Mondrian's own finite-threshold condition e_c<α. The current text warns against conflating them, but a one-sentence numerical illustration would help, since the two thresholds differ by two orders of magnitude at α=0.1, δ=0.1, K=60.
  3. [Checklist item 3(c)] The checklist admirably discloses that the tables print means without error bars, that four groups of printed values lack a retained artifact, and that no code is released. For the empirical part, this limits reproducibility; please make the per-seed data available at least for the headline recovery numbers.
  4. [Sec. 5, pseudo-label ceiling] Please define 'gain' and 'tail' precisely when reporting the PPI++ ceiling ('median gain is modest and only a small tail reaches three times the baseline'), and give the number of classes and the coverage-collapse threshold used.
  5. [General] Several references are to 2026 preprints; please check that they are publicly available or replace them with peer-reviewed versions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: theoretical results are standalone bounds and the empirical recovery is explicitly post hoc.

full rationale

The paper's derivation chain does not reduce to its own inputs. Proposition 1 is the standard Mondrian conformal guarantee, cited to external prior work (Vovk et al., Löfström et al., Ding et al.) and proved via within-class exchangeability; it is not fitted to target data. Proposition 2 is an indistinguishability construction: it exhibits two target joint laws sharing the same P_S and P_T(X) but differing in class-c quantiles, so the identical output-distribution argument is a standard impossibility argument rather than a restatement of the conclusion. Theorem 1's upper bound is a PAC order-statistic bound with a union bound, and its lower bounds are existential hard instances whose rates match the upper bound; neither side is calibrated to the other. The empirical sections explicitly label the recovery pattern as 'an after the fact pattern that no threshold, prospective test, or theorem yet turns into a validated decision rule,' and the pseudo-label experiment is presented as an oracle ceiling with oracle-tuned λ, i.e., an upper bound on a control span rather than a fitted predictor of the observed recovery. There are no self-citations by the authors, so no self-citation chain is load-bearing. Deferred proofs in the supplementary material are a verification gap, not a circular step. Accordingly, the paper receives a circularity score of 0.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The main theorems introduce no free parameters: α, δ, and ε are user-set risk/tolerance levels, and the order-statistic construction is parameter-free. The empirical pseudo-label ceiling uses an oracle-tuned PPI λ and the weighted-conformal baseline uses discriminator constants deferred to the missing supplement, so these are listed below. The paper's load-bearing assumptions are exchangeability (for the Mondrian guarantee), a positive density lower bound (for the sufficiency rate), score separation and joint-shift structure (for the impossibility construction), a known target covariate marginal, and the adequacy of the SPD representation for the case study. No new physical or model entities are introduced.

free parameters (2)
  • PPI control-variate weight λ (oracle-tuned) = oracle (best-case over λ; unbounded unlabeled pool; true target quantile)
    Section 5's pseudo-label ceiling is computed with the best possible λ to bound every deployable PPI estimator in the evaluated span; it is an upper-bound tuning parameter, not an estimate, and its choice affects the reported 1.08× gain.
  • Weighted-conformal density-ratio discriminator constants = unreported in main text (supplementary material, not included)
    Section 3 uses a regularized logistic discriminator separating source from target on a held-out pool and defers the constants to the supplementary material; these hand-chosen values affect the weighted-conformal baseline results.
assumptions (6)
  • domain assumption Within-class exchangeability of source calibration scores and target test scores for every class c
    Used in Proposition 1 and by the empirical source Mondrian remedy; Remark 1 admits deployment shift violates it, so the label-free recovery result is approximate and would fail if within-class score laws shift.
  • domain assumption Target class-conditional score distribution F_T^c is continuous and has density ≥ f0_c > 0 on [q*_c, q*_c + r_c]
    Theorem 1(b) needs this for the sufficiency rate; without a positive density the order-statistic quantile recovery bound (and the abstract's Θ(ε^{-2} logK) rate as stated) need not hold.
  • domain assumption Score separation condition for the class in Proposition 2: disjoint events A,B with P_T(A),P_T(B)>0 and separated score ranges
    The two-source construction requires such separation; it delimits the regime where the target class-conditional quantile is unidentified.
  • domain assumption The shift acts jointly on covariates and labels, i.e., P_T(Y|X) is unrestricted relative to P_S
    The impossibility and label complexity results are for joint shift; if the shift is only covariate or label shift, weighted/label-shift methods may restore guarantees, as the paper notes in §2/§4.
  • domain assumption The unlabeled target covariate marginal P_T(X) is available (or estimable) at deployment
    Both the label-free impossibility and the unlabeled-pool pseudo-label experiments assume a target covariate sample exists; the theorem says this pool cannot replace labeled target pairs.
  • domain assumption SPD descriptor / log-Euclidean geometry is a sufficient representation for the skeleton-action case study
    The empirical recovery claim is instantiated on temporal-covariance SPD descriptors; the paper acknowledges the backbone lags stronger heads, so the representational adequacy is load-bearing for the case study.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Label Complexity of Class-Conditional Coverage under Distribution Shift." pith.science (2026). https://pith.science/paper/QAHZQCXH

@misc{pith2026260718088,
  author       = {Pith},
  title        = {Pith review of: The Label Complexity of Class-Conditional Coverage under Distribution Shift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QAHZQCXH}},
  note         = {Machine review of arXiv:2607.18088}
}
read the original abstract

Conformal prediction certifies that a classifier's prediction sets cover the truth, and that certificate is marginal. Many recognition benchmarks build distribution shift into evaluation, placing disjoint conditions in the training and test splits. Under that shift the certificate stays reassuring while per class coverage fails silently: on a real cross subject skeleton benchmark marginal coverage holds near ninety percent while the worst class is covered about seventy percent and ten of sixty classes fall below eighty percent. This class specific undercoverage stays hidden behind a single reassuring marginal number. Once the shift acts jointly on covariates and labels, the target class conditional score law is unidentified, so no label free method is at once per class valid and efficient uniformly over target laws consistent with the observed source joint distribution and target covariate marginal. The per class labels needed to recover every class threshold to a given tolerance grow as the inverse square of that tolerance and the logarithm of the class count, with matching bounds for classwise threshold procedures. Pseudo labels do not shortcut it: the best prediction powered estimator gains at most a small constant factor where coverage collapses. Across three real shifts and an image corruption benchmark, source label calibration recovers much of the gap while marginal coverage holds, and stops once it breaks.

Figures

Figures reproduced from arXiv: 2607.18088 by the authors.

Figure 1
Figure 1. Marginal coverage hides a per-class collapse; among the methods we evaluate, only target labels restore it. (1) On NTU-60 cross-subject skeleton action recognition, split conformal holds marginal coverage at 0.881; per-class coverage collapses at the same time. The worst class reaches 0.699 and 10 of 60 classes fall below 0.80. (2) In the validity (worst-class coverage) versus efficiency (mean set size) plane, every… view at source ↗
Figure 2
Figure 2. Cross subject shift: fixed calibration, uneven per class drift (schematic). A fixed source calibrated region undercovers the classes whose per class score mass drifts out of it at deployment on disjoint subjects; the annotated numbers are the real NTU-60 split conformal result. Action recognition: accuracy versus reliabil￾ity. Conformal prediction has been applied to hu￾man action recognition using vision language m… view at source ↗
Figure 3
Figure 3. Label free signals on the SPD mani￾fold (synthetic schematic). A target point drifts toward the wrong class mean M3. The minimum dis￾tance signal calls it in support and misses the mis￾coverage (AUROC 0.55); the geodesic margin sees d(x, M1) ≈ d(x, M3) and flags it (AUROC 0.89). Both AUROC values are measured on the synthetic study; the geometry is illustrative. Neither signal improves worst class coverage on real d… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: [Real data, NTU-60 cross subject] Per class coverage recovery on the main shift. (a) Worst class coverage against mean set size (dashed line: non dominated set): source Mondrian sits between split conformal and the oracle, closing much of the gap (0.70 to 0.77) near no…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

72 extracted references · 4 linked inside Pith

  1. [1]

    Angelman, S., Nizhar, R., and Goldberger, J. (2025). Calibrating without labels: Source-free conformal prediction using pseudo-labels. In Symposium on Conformal and Probabilistic Prediction with Applications (COPA) , pages 63--81. PMLR

    2025

  2. [2]

    N., Bates, S., Fannjiang, C., Jordan, M

    Angelopoulos, A. N., Bates, S., Fannjiang, C., Jordan, M. I., and Zrnic, T. (2023a). Prediction-powered inference. Science , 382(6671):669--674

  3. [3]

    N., Bates, S., Fisch, A., Lei, L., and Schuster, T

    Angelopoulos, A. N., Bates, S., Fisch, A., Lei, L., and Schuster, T. (2024). Conformal risk control. In International Conference on Learning Representations (ICLR)

    2024

  4. [4]

    N., Bates, S., Malik, J., and Jordan, M

    Angelopoulos, A. N., Bates, S., Malik, J., and Jordan, M. I. (2021). Uncertainty sets for image classifiers using conformal prediction. In International Conference on Learning Representations (ICLR)

    2021

  5. [6]

    Arsigny, V., Fillard, P., Pennec, X., and Ayache, N. (2006). Log-euclidean metrics for fast and simple calculus on diffusion tensors. Magnetic Resonance in Medicine , 56(2)

    2006

  6. [7]

    F., Cand \`e s, E

    Barber, R. F., Cand \`e s, E. J., Ramdas, A., and Tibshirani, R. J. (2023). Conformal prediction beyond exchangeability. The Annals of Statistics , 51(2)

    2023

  7. [8]

    Bary, T., Fuchs, C., and Macq, B. (2025). Conformal predictions for human action recognition with vision-language models. In IEEE International Conference on Image Processing Workshops (ICIPW)

    2025

  8. [9]

    Bhatia, R. (2007). Positive Definite Matrices . Princeton University Press

    2007

Show all 72 references
  1. [11]

    Chen, Y., Zhang, Z., Yuan, C., Li, B., Deng, Y., and Hu, W. (2021). Channel-wise topology refinement graph convolution for skeleton-based action recognition. In IEEE/CVF International Conference on Computer Vision (ICCV)

    2021

  2. [12]

    N., Bates, S., Jordan, M

    Ding, T., Angelopoulos, A. N., Bates, S., Jordan, M. I., and Tibshirani, R. J. (2023). Class-conditional conformal prediction with many classes. In Advances in Neural Information Processing Systems , volume 36. arXiv:2306.09335

    2023 arXiv

  3. [13]

    and Cand \`e s, E

    Gibbs, I. and Cand \`e s, E. J. (2021). Adaptive conformal inference under distribution shift. In Advances in Neural Information Processing Systems (NeurIPS)

    2021

  4. [14]

    and Van Gool, L

    Huang, Z. and Van Gool, L. (2017). A riemannian network for SPD matrix learning. In AAAI Conference on Artificial Intelligence

    2017

  5. [15]

    Kasa, K., Zhang, Z., Yang, H., and Taylor, G. W. (2025). Adapting prediction sets to distribution shifts without labels. In Uncertainty in Artificial Intelligence (UAI) , pages 1990--2010. PMLR

    2025

  6. [16]

    and Zhou, J

    Khanal, A. and Zhou, J. (2026). Severe domain shift in skeleton-based action recognition: A study of uncertainty failure in real-world gym environments. arXiv preprint arXiv:2603.15574 . Preprint

    2026

  7. [17]

    J., and Wasserman, L

    Lei, J., G'Sell, M., Rinaldo, A., Tibshirani, R. J., and Wasserman, L. (2018). Distribution-free predictive inference for regression. Journal of the American Statistical Association , 113(523)

    2018

  8. [18]

    o fstr \

    L \"o fstr \"o m, T., Bostr \"o m, H., Linusson, H., and Johansson, U. (2015). Bias reduction through conditional conformal prediction. Intelligent Data Analysis , 19(6):1355--1375

    2015

  9. [19]

    Pennec, X. (2006). Intrinsic statistics on riemannian manifolds: Basic tools for geometric measurements. Journal of Mathematical Imaging and Vision , 25(1)

    2006

  10. [20]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. (2021). Learning transferable visual models from natural language supervision. In International Conference on Machine Learning...

    2021

  11. [21]

    Romano, Y., Sesia, M., and Cand \`e s, E. J. (2020). Classification with valid and adaptive coverage. In Advances in Neural Information Processing Systems (NeurIPS)

    2020

  12. [22]

    Sadinle, M., Lei, J., and Wasserman, L. (2019). Least ambiguous set-valued classifiers with bounded error levels. Journal of the American Statistical Association , 114(525)

    2019

  13. [23]

    Shahroudy, A., Liu, J., Ng, T.-T., and Wang, G. (2016). NTU RGB+D : A large scale dataset for 3D human activity analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2016

  14. [25]

    J., Barber, R

    Tibshirani, R. J., Barber, R. F., Cand \`e s, E. J., and Ramdas, A. (2019). Conformal prediction under covariate shift. In Advances in Neural Information Processing Systems (NeurIPS)

    2019

  15. [26]

    Vovk, V., Gammerman, A., and Shafer, G. (2005). Algorithmic Learning in a Random World . Springer

    2005

  16. [27]

    Vovk, V., Lindsay, D., Nouretdinov, I., and Gammerman, A. (2003). Mondrian confidence machine. On-line Compression Modelling Project (Working Paper) 4, Computer Learning Research Centre, Royal Holloway, University of London

    2003

  17. [28]

    and Qiao, X

    Wang, B. and Qiao, X. (2025). Conformal prediction under generalized covariate shift with posterior drift. In International Conference on Artificial Intelligence and Statistics (AISTATS) , pages 4888--4896. PMLR

    2025

  18. [29]

    Wang, J., Nie, X., Xia, Y., Wu, Y., and Zhu, S.-C. (2014). Cross-view action modeling, learning, and recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2014

  19. [30]

    Yan, S., Xiong, Y., and Lin, D. (2018). Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI Conference on Artificial Intelligence

    2018

  20. [31]

    Algorithmic Learning in a Random World , author=

  21. [32]

    Journal of the American Statistical Association , volume=

    Distribution-Free Predictive Inference for Regression , author=. Journal of the American Statistical Association , volume=

  22. [33]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Conformal Prediction Under Covariate Shift , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  23. [34]

    The Annals of Statistics , volume=

    Conformal Prediction Beyond Exchangeability , author=. The Annals of Statistics , volume=

  24. [35]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Adaptive Conformal Inference Under Distribution Shift , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  25. [36]

    Journal of the American Statistical Association , volume=

    Least Ambiguous Set-Valued Classifiers With Bounded Error Levels , author=. Journal of the American Statistical Association , volume=

  26. [37]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Classification with Valid and Adaptive Coverage , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  27. [38]

    International Conference on Learning Representations (ICLR) , year=

    Uncertainty Sets for Image Classifiers Using Conformal Prediction , author=. International Conference on Learning Representations (ICLR) , year=

  28. [39]

    International Conference on Learning Representations (ICLR) , year=

    Conformal Risk Control , author=. International Conference on Learning Representations (ICLR) , year=

  29. [40]

    arXiv preprint arXiv:2507.04441 , year=

    A Category-Theoretic Analysis of Conformal Prediction , author=. arXiv preprint arXiv:2507.04441 , year=

    arXiv

  30. [41]

    The Annals of Statistics , volume=

    Convergence of Estimates Under Dimensionality Restrictions , author=. The Annals of Statistics , volume=

  31. [42]

    Introduction to Nonparametric Estimation , author=

  32. [43]

    Festschrift for Lucien Le Cam , editor=

    Assouad, Fano, and Le Cam , author=. Festschrift for Lucien Le Cam , editor=

  33. [44]

    Science , volume=

    Prediction-powered inference , author=. Science , volume=

  34. [45]

    arXiv preprint arXiv:2311.01453 , year=

    PPI++: Efficient Prediction-Powered Inference , author=. arXiv preprint arXiv:2311.01453 , year=

    arXiv

  35. [46]

    A Riemannian Network for

    Huang, Zhiwu and Van Gool, Luc , booktitle=. A Riemannian Network for

  36. [47]

    Journal of Mathematical Imaging and Vision , volume=

    Intrinsic Statistics on Riemannian Manifolds: Basic Tools for Geometric Measurements , author=. Journal of Mathematical Imaging and Vision , volume=

  37. [48]

    Positive Definite Matrices , author=

  38. [49]

    Magnetic Resonance in Medicine , volume=

    Log-Euclidean Metrics for Fast and Simple Calculus on Diffusion Tensors , author=. Magnetic Resonance in Medicine , volume=

  39. [50]

    AAAI Conference on Artificial Intelligence , year=

    Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition , author=. AAAI Conference on Artificial Intelligence , year=

  40. [51]

    IEEE/CVF International Conference on Computer Vision (ICCV) , year=

    Channel-Wise Topology Refinement Graph Convolution for Skeleton-Based Action Recognition , author=. IEEE/CVF International Conference on Computer Vision (ICCV) , year=

  41. [52]

    Shahroudy, Amir and Liu, Jun and Ng, Tian-Tsong and Wang, Gang , booktitle=

  42. [53]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Cross-View Action Modeling, Learning, and Recognition , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  43. [54]

    IEEE International Conference on Image Processing Workshops (ICIPW) , year=

    Conformal Predictions for Human Action Recognition with Vision-Language Models , author=. IEEE International Conference on Image Processing Workshops (ICIPW) , year=

  44. [55]

    Severe Domain Shift in Skeleton-Based Action Recognition:

    Khanal, Aaditya and Zhou, Junxiu , journal=. Severe Domain Shift in Skeleton-Based Action Recognition:. 2026 , note=

    2026

  45. [56]

    2003 , institution=

    Mondrian Confidence Machine , author=. 2003 , institution=

    2003

  46. [57]

    Advances in Neural Information Processing Systems , volume=

    Class-Conditional Conformal Prediction with Many Classes , author=. Advances in Neural Information Processing Systems , volume=. 2023 , note=

    2023

  47. [58]

    International Conference on Machine Learning (ICML) , year=

    Learning Transferable Visual Models From Natural Language Supervision , author=. International Conference on Machine Learning (ICML) , year=

  48. [59]

    Proceedings of the 42nd International Conference on Machine Learning , series=

    Kandinsky Conformal Prediction: Beyond Class- and Covariate-Conditional Coverage , author=. Proceedings of the 42nd International Conference on Machine Learning , series=. 2025 , publisher=

    2025

  49. [60]

    Intelligent Data Analysis , volume=

    Bias reduction through conditional conformal prediction , author=. Intelligent Data Analysis , volume=. 2015 , publisher=

    2015

  50. [61]

    Asian Conference on Machine Learning (ACML) , pages =

    Conditional validity of inductive conformal predictors , author =. Asian Conference on Machine Learning (ACML) , pages =. 2012 , organization =

    2012

  51. [62]

    Journal of the Royal Statistical Society: Series B (Statistical Methodology) , volume =

    Distribution-free prediction bands for non-parametric regression , author =. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , volume =

  52. [63]

    Information and Inference: A Journal of the IMA , volume =

    The limits of distribution-free conditional predictive inference , author =. Information and Inference: A Journal of the IMA , volume =

  53. [64]

    Journal of the American Statistical Association , volume =

    Robust validation: Confident predictions even when distributions shift , author =. Journal of the American Statistical Association , volume =

  54. [65]

    Uncertainty in Artificial Intelligence (UAI) , pages =

    Distribution-free uncertainty quantification for classification under label shift , author =. Uncertainty in Artificial Intelligence (UAI) , pages =. 2021 , organization =

    2021

  55. [66]

    Uncertainty in Artificial Intelligence (UAI) , pages =

    Adapting Prediction Sets to Distribution Shifts Without Labels , author =. Uncertainty in Artificial Intelligence (UAI) , pages =. 2025 , organization =

    2025

  56. [67]

    Symposium on Conformal and Probabilistic Prediction with Applications (COPA) , pages =

    Calibrating Without Labels: Source-Free Conformal Prediction Using Pseudo-Labels , author =. Symposium on Conformal and Probabilistic Prediction with Applications (COPA) , pages =. 2025 , organization =

    2025

  57. [68]

    International Conference on Artificial Intelligence and Statistics (AISTATS) , pages =

    Conformal Prediction Under Generalized Covariate Shift with Posterior Drift , author =. International Conference on Artificial Intelligence and Statistics (AISTATS) , pages =. 2025 , organization =

    2025

  58. [69]

    arXiv preprint arXiv:2602.14913 , year =

    Coverage Guarantees for Pseudo-Calibrated Conformal Prediction under Distribution Shift , author =. arXiv preprint arXiv:2602.14913 , year =

    arXiv

  59. [70]

    Bairaktari, K., Wu, J., and Wu, Z. S. (2025). Kandinsky conformal prediction: Beyond class- and covariate-conditional coverage. In Proceedings of the 42nd International Conference on Machine Learning , volume 267 of Proceedings of Machine Learning Research , pages 2581--2602. PMLR

    2025

  60. [71]

    F., Cand \`e s, E

    Barber, R. F., Cand \`e s, E. J., Ramdas, A., and Tibshirani, R. J. (2021). The limits of distribution-free conditional predictive inference. Information and Inference: A Journal of the IMA , 10(2):455--482

    2021

  61. [72]

    Cauchois, M., Gupta, S., Ali, A., and Duchi, J. C. (2024). Robust validation: Confident predictions even when distributions shift. Journal of the American Statistical Association , 119(548):3033--3044

    2024

  62. [73]

    and Wasserman, L

    Lei, J. and Wasserman, L. (2014). Distribution-free prediction bands for non-parametric regression. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 76(1):71--96

    2014

  63. [74]

    and Ramdas, A

    Podkopaev, A. and Ramdas, A. (2021). Distribution-free uncertainty quantification for classification under label shift. In Uncertainty in Artificial Intelligence (UAI) , pages 844--853. PMLR

    2021

  64. [75]

    Vovk, V. (2012). Conditional validity of inductive conformal predictors. In Asian Conference on Machine Learning (ACML) , pages 475--490. PMLR

    2012

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.