Pith. sign in

REVIEW 3 major objections 5 minor 27 references

Wasserstein-regularized Conformal Prediction under General Distribution Shift

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A Wasserstein-based bound splits conformal prediction's coverage gap into covariate- and concept-shift terms, yielding a training objective that keeps gaps near 3.2% while cutting prediction set size by 37% relative to worst-case baselines.

desk verdict Wasserstein bound for conformal coverage is a useful idea, but the main theorem silently drops the label noise that the paper's own model assumes, so Eq. (13) does not hold as stated. read the letter →

arxiv 2501.13430 v2 pith:NBNIP3BB submitted 2025-01-23 cs.LG stat.ML

Rui Xu , Chao Chen , Yue Sun , Parvathinathan Venkitasubramaniam , Sihong Xie This is my paper · ORCID
classification cs.LGstat.ML
keywords conformalpredictiondistributionshiftWassersteindistancecoveragegapcovariateconceptimportanceweightingrepresentationlearning
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 claims that the coverage gap of conformal prediction under a joint distribution shift is governed by the Wasserstein distance between calibration and test conformal scores, not by total variation distance. It proves a bound of the form coverage gap ≤ $\sqrt{2L(\kappa W(P_X,Q_X)+\eta W(Q_{Y,f_P},Q_Y))}$, separating the effect of covariate shift from concept shift. On this decomposition it builds WR-CP, an algorithm that combines importance weighting with a Wasserstein-regularized training objective to shrink both terms. Experiments on six datasets report coverage gaps reduced to 3.2% across confidence levels and prediction sets 37% smaller than the worst-case approach on average. If correct, this gives practitioners a computable, optimizable handle on conformal accuracy under shifts that are neither purely covariate nor purely label-based.

What carries the argument

The load-bearing mechanism is the triple of pushforward measures $P_V = s_P\#P_X$, $Q_{V,s_P} = s_P\#Q_X$, and $Q_V = s_Q\#Q_X$ on conformal score space, together with the triangle inequality $W(P_V,Q_V) \le W(P_V,Q_{V,s_P}) + W(Q_{V,s_P},Q_V)$. The first term is bounded by $\kappa W(P_X,Q_X)$ via a Lipschitz contraction (Theorem 2), and the second is bounded by $\eta W(Q_{Y,f_P},Q_Y)$ via a ratio bound that compares score differences to label-function differences. Importance weighting converts $P_V$ toward $Q_{V,s_P}$, which handles covariate shift; the WR-CP objective then minimizes an empirical surrogate of the remaining concept-shift term, $\min_\theta \sum_i \mathbb{E}_{D^{(i)}_{XY}}[\ell(h_\theta(x),y)] + \beta \sum_i W(\hat{D}^{(i)}_{V,s_P}, \hat{D}^{(i)}_V)$, with kernel density estimation for likelihood ratios and point-wise CDF estimation to keep the Wasserstein term differentiable. A finite-sample bound from empirical Wasserstein convergence (Weed & Bach, 2019) turns the population bound into a usable guarantee with explicit sample-size terms.

What would settle it

Construct calibration and test data with identical covariate distributions and the same true regression function $f$, but different label noise variances, e.g., $y|x \sim N(f(x), 0.1)$ for calibration and $y|x \sim N(f(x), 10)$ for test; the paper's bound in Eq. (13) then predicts a zero coverage gap, while split conformal prediction on test data will show coverage falling well below $1-\alpha$.

Watch

Extended reading notes

Core claim

The central claim is that the absolute difference between the calibration and test conformal CDFs at the chosen quantile is bounded by $\sqrt{2L(\kappa W(P_X,Q_X)+\eta W(Q_{Y,f_P},Q_Y))}$, where $L$ bounds the calibration score density, $\kappa$ is the Lipschitz constant of the score function, and $\eta$ is the maximal ratio of score differences to label-function differences. The paper obtains this by pushing the calibration and test distributions through the score function to conformal score space, then using the triangle inequality to isolate a covariate-shift term and a concept-shift term. It shows that importance weighting eliminates the covariate-shift term even when concept shift is present, and that the concept-shift term can be minimized through a regularization term during model training. The resulting WR-CP method is evaluated on multi-source domain generalization, where the test distribution is an unknown mixture of training distributions, and is reported to keep coverage close to the target while reducing prediction set size substantially.

Load-bearing premise

The proof requires the conformal score at an input $x$ to be a deterministic function of $x$, so any shift in label noise between calibration and test distributions is invisible to the bound and to WR-CP's training objective.

Editorial extensions

If this is right

  • Practitioners can use the estimated Wasserstein distance between calibration and test scores as an early-warning signal for coverage violations at any specific confidence level $\alpha$.
  • Because the decomposition isolates covariate and concept shifts, importance weighting can be applied safely under joint shift without losing its covariate-shift correction.
  • The dependence of the bound on $\kappa$ and $\eta$ shows that model training can directly shrink the coverage gap by promoting smooth, shift-insensitive residual functions.
  • The tunable parameter $\beta$ yields a Pareto front between coverage accuracy and prediction set efficiency, so WR-CP can be adapted to application-specific risk preferences.
  • In multi-source settings, the bound implies that controlling the largest per-source weighted gap suffices to guarantee coverage on any unknown mixture of the source distributions.

Reading between the lines

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

  • A direct extension would replace the deterministic score representation with a stochastic one that includes label noise; the bound would then acquire an additional term proportional to the Wasserstein distance between the noise distributions of calibration and test, which the current decomposition misses.
  • The same pushforward decomposition can be coupled with direct optimal-transport estimators of $W(P_X,Q_X)$ and $W(Q_{Y,f_P},Q_Y)$, which may be more robust than the kernel-density-based proxy used for likelihood ratios.
  • The empirical bound in Eq. (16) suggests a practical rule: inflate the conformal threshold by a margin proportional to the empirical Wasserstein distance plus sample-size terms, yielding a finite-sample coverage guarantee that WR-CP can then minimize.
  • Because the method depends on kernel density estimation for importance weights, its behavior in high-dimensional feature spaces is an open question and is likely to require density-ratio estimators that scale better.
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 manuscript studies conformal prediction under joint covariate and concept shift. It proposes bounding the coverage gap by the Wasserstein distance between calibration and test conformal score distributions, decomposes that distance into covariate- and concept-shift terms via pushforward measures, and introduces WR-CP, an algorithm that combines importance weighting with a Wasserstein regularizer on the score distributions. Theoretical statements include a bound of the coverage gap in Eq. (13), a finite-sample empirical bound in Eq. (16), and an application to multi-source domain generalization. Experiments on six datasets compare WR-CP with vanilla CP, IW-CP, CQR, and worst-case CP, reporting reduced coverage gaps (to 3.2%) and smaller prediction sets (37%).

Significance. The high-level idea is appealing and the empirical study is substantial; if the theoretical claims were fully supported, the paper would make a useful contribution to conformal prediction under distribution shift. Strengths include a clearly stated decomposition of the Wasserstein bound, a released implementation, extensive multi-trial experiments across six datasets, and an explicit discussion of limitations in Appendix H. The main reservation is that the central bound as stated does not cover the label-noise component of the score distribution, so the theoretical support for the algorithm is incomplete.

major comments (3)
  1. [Section 3.2, Eq. (7) and Eq. (13)] The paper defines PV=sP#PX and QV=sQ#QX with sP(x)=|h(x)-fP(x)| and sQ(x)=|h(x)-fQ(x)|, after Section 1 assumes y|x ~ N(fP(x), epsP) and y|x ~ N(fQ(x), epsQ). The actual conformal scores are |h(x)-y|, so their distributions are convolutions of these deterministic residuals with label noise. A pure noise shift with PX=QX and fP=fQ but epsP != epsQ makes both Wasserstein terms in Eq. (13) vanish, while the true coverage gap can be large; for example, with alpha=0.2, epsP=1, epsQ=2, and Gaussian noise, the calibration quantile is about 1.28 and the test coverage is about 0.48. Thus Eq. (13) does not bound the actual coverage gap unless one adds an explicit assumption such as epsP=epsQ or epsP=epsQ=0, which is not stated in the paper. Since Algorithm 1 lines 3-7 use the actual noisy scores, the object being optimized is not the noiseless surrogate analyzed by the theory.
  2. [Eq. (10) and definition of eta] The constant eta = max_{x1,x2} |sP(x1)-sQ(x2)| / |fP(x1)-fQ(x2)| is not guaranteed to be finite. If fP(x1)=fQ(x2) for some pair with a nonzero numerator, or if fP=fQ on a set of positive measure, the quotient is either undefined or infinite, while the denominator term W(QY,fP,QY) may be zero. The inequality in Eq. (10) therefore requires a condition such as |fP(x1)-fQ(x2)| >= c > 0 on the support of the transport plan, or an alternative Lipschitz-style control of |sP-sQ| in terms of |fP-fQ|. Without such a condition, the bound in Eq. (13) is not well posed.
  3. [Section 5, Eq. (19)] The derivation of the multi-source surrogate is incomplete. Theorem 4 yields W(QV,sP,QV) <= sum_i w_i sum_j w_j W(DV,sP^j, DV^i), a double sum over all pairs of source domains. With uniform weights this is (1/k^2) sum_{i,j} W(DV,sP^j, DV^i), but the objective minimized in Eq. (19) is (1/k) sum_i W(DV,sP^i, DV^i), which is only the diagonal part. No stated inequality shows that the diagonal sum controls the double sum. The paper should either prove such an inequality, add additional assumptions on the source distributions, or explicitly state that Eq. (19) is a heuristic surrogate rather than an upper bound on the concept-shift term.
minor comments (5)
  1. [Definition 3] There is a typo: 'prbability' should be 'probability'.
  2. [Definition 4] There is a typo: 'Wassersteion dimension' should be 'Wasserstein dimension'.
  3. [Appendix A.4] The phrase 'definded on X' should be 'defined on X'.
  4. [Appendix F.1] The dataset is referred to as 'PeMSED8' in one place, while the rest of the paper consistently uses 'PeMSD8'.
  5. [Proof of Theorem 3] The proof says it is 'reasonable to assume' that the two concentration events are independent; this should be stated as a consequence of the empirical measures being built from independent samples rather than as an additional assumption.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the Wasserstein coverage-gap bound is derived from explicit transport inequalities, and the reported coverage gaps are measured on held-out test data rather than read off from the minimized surrogate.

full rationale

Walk of the derivation: Eq. (3) defines the coverage gap as |F_PV(tau) - F_QV(tau)|; Eq. (4) bounds any pointwise CDF gap by the Kolmogorov distance; Eq. (6) applies Ross's bound K <= sqrt(2L W); Eq. (7) is the triangle inequality on pushforward measures; Eqs. (10)-(13) combine the eta-ratio bound and the Lipschitz bound with Theorem 2. Every step is an explicit inequality over distributions that are defined independently of the reported test set. Importance weighting is used as a measure change (dQX/dPX) so that weighted PV becomes QV,sP; this is an exact transformation under the likelihood-ratio model, and the resulting coverage gap is then measured on held-out test data, not read off from the minimized Wasserstein term. Beta is a tunable hyperparameter, and performances are benchmarked against vanilla CP, IW-CP, CQR, and WC-CP, so the empirical claims are not forced by the fitted parameters. The related-work citation to Xu and Xie (2021) is not load-bearing, and no uniqueness theorem is imported from the authors' prior work. The serious caveat in the paper is a correctness/assumption gap rather than circularity: Section 3.2 defines PV = sP#PX and QV = sQ#QX, treating scores as deterministic functions of the covariate, although Section 1 models y|x as N(fP(x), epsP) and N(fQ(x), epsQ); a pure change in label noise epsP != epsQ would make both Wasserstein terms in Eq. (13) vanish while the true score distributions differ. This means the theorem is not established for the noisy scores used in Algorithm 1, but it does not make the bound equal to its own input. Appendices H.1-H.3 and G candidly state estimation errors, computational costs, calibration choices, and datasets where efficiency gains do not materialize; none of these admissions creates a circular derivation. No step renames a known result as a new prediction or derives the target conclusion from itself.

Assumptions & free parameters 2 free parameters · 7 assumptions · 0 invented entities

The ledger shows the central bound rests on standard optimal transport inequalities plus strong domain assumptions: noiseless or identical-noise score pushforwards, mixture structure of test distribution, and finite eta. No new physical entities are introduced.

free parameters (2)
  • beta (WR-CP regularization weight) = dataset- and alpha-dependent; e.g., 4.5, 9, 9, 6, 8, 20 for alpha=0.2
    Controls trade-off between coverage gap and set size; tuned per dataset and alpha, listed in Figure 4 and Table 2.
  • KDE bandwidth = grid-searched per dataset (log range 1e-2 to 10^0.5, 20 values)
    Density estimates drive likelihood ratios; bandwidth chosen by grid search (Appendix E.1 and H.1).
assumptions (7)
  • domain assumption y|x is Gaussian: y|x ~ N(fP(x), epsP) and N(fQ(x), epsQ)
    Stated in Section 1 to model concept shift; epsP and epsQ are never estimated and are not used later.
  • ad hoc to paper Conformal score is a deterministic function of x: PV=sP#PX and QV=sQ#QX with sP(x)=|h(x)-fP(x)|
    Section 3.2 uses this to build the decomposition; it conflicts with the Gaussian label noise assumption and omits noise shift.
  • domain assumption Test distribution is an unknown mixture of k training distributions (Eq. 17)
    Defines the multi-source domain generalization setup in Section 5.
  • domain assumption Calibration distribution is the uniform mixture of training distributions
    Used in experiments (Section 6.1); Appendix H.3 notes this may not hold in general.
  • ad hoc to paper eta = sup |sP(x1)-sQ(x2)| / |fP(x1)-fQ(x2)| is finite
    Eq. (10) requires the denominator to be nonzero wherever the numerator is nonzero; this is not stated or verified.
  • standard math Density bound L for score distributions exists
    Needed for Proposition 1; assumed for PV and QV.
  • standard math Wasserstein-1 on R equals area between CDFs; triangle inequality and pushforward properties
    Used in Theorems 2, 3, and 4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Wasserstein-regularized Conformal Prediction under General Distribution Shift." pith.science (2026). https://pith.science/paper/NBNIP3BB

@misc{pith2026250113430,
  author       = {Pith},
  title        = {Pith review of: Wasserstein-regularized Conformal Prediction under General Distribution Shift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NBNIP3BB}},
  note         = {Machine review of arXiv:2501.13430}
}
abstract

Conformal prediction yields a prediction set with guaranteed $1-\alpha$ coverage of the true target under the i.i.d. assumption, which may not hold and lead to a gap between $1-\alpha$ and the actual coverage. Prior studies bound the gap using total variation distance, which cannot identify the gap changes under distribution shift at a given $\alpha$. Besides, existing methods are mostly limited to covariate shift,while general joint distribution shifts are more common in practice but less researched.In response, we first propose a Wasserstein distance-based upper bound of the coverage gap and analyze the bound using probability measure pushforwards between the shifted joint data and conformal score distributions, enabling a separation of the effect of covariate and concept shifts over the coverage gap. We exploit the separation to design an algorithm based on importance weighting and regularized representation learning (WR-CP) to reduce the Wasserstein bound with a finite-sample error bound.WR-CP achieves a controllable balance between conformal prediction accuracy and efficiency. Experiments on six datasets prove that WR-CP can reduce coverage gaps to $3.2\%$ across different confidence levels and outputs prediction sets 37$\%$ smaller than the worst-case approach on average.

Figures

Figures reproduced from arXiv: 2501.13430 by the authors.

Figure 1
Figure 1. (a) Joint distribution shift can include both covariate shift ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Pushforward measures. With Eq. (7) bounding W(PV , QV ), we design an approach to minimize the upper bound. First, we adopt importance weighting, which weights calibration conformal scores with the likelihood ratio dQX(x)/dPX(x). Tibshirani et al. (2019) prove that impor￾tance weighting can preserve the coverage guarantee when only a covariate shift occurs. However, existing works do not include the weighting techni… view at source ↗
Figure 3
Figure 3. Comparison of vanilla CP, IW-CP, and WR-CP based on normalized Wasserstein distance between calibration and test conformal scores: IW-CP can only address the distance caused by covariate shift, while WR-CP reduces the distance from concept shift. The β values for the WR-CP method are 9, 11, 9, 10, 13, and 13, respectively. 6.4 ROBUST AND EFFICIENT PREDICTION SETS BY WR-CP We experimentally demonstrated that, compare… view at source ↗
Figures from the paper (18 more)
Figure 4
Figure 4. Figure 4: Coverages and prediction set sizes of WR-CP and baselines with α = 0.2: WR-CP makes coverages on test data more concentrated around the 1 − α level compared to vanilla CP, IW-CP, and CQR. While WC-CP ensures coverage guarantees, it leads to inefficient predictions due …
Figure 5
Figure 5. Figure 5: Pareto fronts of coverage gap and prediction set size obtained from WR-CP with varying β: WR-CP effectively balances conformal prediction accuracy and efficiency, providing a flexible and customizable solution. When β = 0, WR-CP returns to IW-CP. like vanilla CP and IW…
Figure 6
Figure 6. Figure 6: Comparison between total variation distance and Wasserstein distance: a reduction in the total variation distance does not necessarily result in CDFs becoming closer. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 3
Figure 3. Figure 3: This phenomenon is further illustrated in Figure 7(b). [PITH_FULL_IMAGE:figures/full_fig_p017_3.png]
Figure 7
Figure 7. Figure 7: (a) and evidenced by the results for the airfoil self-noise, PeMSD4, PeMSD8, U.S.-States, and Japan-Prefectures datasets in [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Comparison between WR-CP and WR-CP(uw) at α = 0.2. Both methods were implemented using the same β values of 4.5, 9, 9, 6, 8, and 20 across the datasets. D GEOMETRIC INTUITION OF η To provide a geometric intuition of η, we expand the definition of η as η = max x1,x2∈X |…
Figure 9
Figure 9. Figure 9: Geometric intuition of η when (a) x1 = x2 and (b) x1 ̸= x2: Intuitively, the residual difference caused by concept shift will be constrained by η. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Calibration and test samples are not exchangeable as they are from different distributions. F.2 SPEARMAN’S COEFFICIENT We first provide the definition of Pearson coefficient. Definition 5 (Pearson coefficient). With n pairs of samples, (xi , yi) for i = 1, ..., n, of …
Figure 11
Figure 11. Figure 11: Coverages and set sizes of WR-CP and baselines with α = 0.1: The β values for the WR-CP method are 9, 11, 9, 8, 13, and 20, respectively. 0.4 0.6 0.8 1.0 Coverage Airfoil 0.6 0.8 PeMSD4 0.4 0.6 0.8 1.0 PeMSD8 0.6 0.7 0.8 0.9 Seattle 0.25 0.50 0.75 1.00 US 0.6 0.8 Japa…
Figure 12
Figure 12. Figure 12: Coverages and set sizes of WR-CP and baselines with α = 0.3: The β values for the WR-CP method are 3, 5, 5, 5, 8, and 13, respectively. 0.25 0.50 0.75 1.00 Coverage Airfoil 0.4 0.6 0.8 PeMSD4 0.4 0.6 0.8 1.0 PeMSD8 0.5 0.6 0.7 0.8 Seattle 0.25 0.50 0.75 1.00 US 0.4 0.…
Figure 13
Figure 13. Figure 13: Coverages and set sizes of WR-CP and baselines with α = 0.4: The β values for the WR-CP method are 3, 5, 5, 5, 8, and 13, respectively. 0.25 0.50 0.75 1.00 Coverage Airfoil 0.4 0.6 0.8 PeMSD4 0.25 0.50 0.75 1.00 PeMSD8 0.4 0.5 0.6 0.7 Seattle 0.5 1.0 US 0.2 0.4 0.6 0.…
Figure 14
Figure 14. Figure 14: Coverages and set sizes of WR-CP and baselines with α = 0.5: The β values for the WR-CP method are 3, 5, 3, 5, 8, and 13, respectively. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_14.png]
Figure 15
Figure 15. Figure 15: Coverages and set sizes of WR-CP and baselines with α = 0.6: The β values for the WR-CP method are 3, 5, 3, 5, 8, and 13, respectively. 0.25 0.50 0.75 1.00 Coverage Airfoil 0.2 0.4 0.6 PeMSD4 0.25 0.50 0.75 PeMSD8 0.3 0.4 Seattle 0.0 0.5 1.0 US 0.00 0.25 0.50 0.75 Jap…
Figure 16
Figure 16. Figure 16: Coverages and set sizes of WR-CP and baselines with α = 0.7: The β values for the WR-CP method are 2, 2, 2, 5, 8, and 10, respectively. 0.25 0.50 0.75 Coverage Airfoil 0.2 0.4 PeMSD4 0.25 0.50 0.75 PeMSD8 0.15 0.20 0.25 0.30 Seattle 0.0 0.5 1.0 US 0.0 0.2 0.4 0.6 Japa…
Figure 17
Figure 17. Figure 17: Coverages and set sizes of WR-CP and baselines with α = 0.8: The β values for the WR-CP method are 2, 2, 2, 5, 5, and 10, respectively. 0.0 0.2 0.4 0.6 Coverage Airfoil 0.1 0.2 0.3 PeMSD4 0.0 0.2 0.4 PeMSD8 0.10 0.15 Seattle 0.00 0.25 0.50 0.75 US 0.0 0.2 0.4 0.6 Japa…
Figure 18
Figure 18. Figure 18: Coverages and set sizes of WR-CP and baselines with α = 0.9: The β values for the WR-CP method are 2, 1, 1, 5, 2, and 6, respectively. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_18.png]
Figure 19
Figure 19. Figure 19: Coverages and set sizes of WC-CP and WR-CP with coverage guarantee at 1 − α − αD. The experiment results are depicted in [PITH_FULL_IMAGE:figures/full_fig_p024_19.png]
Figure 20
Figure 20. Figure 20: Coverages and set sizes of WC-CP and Hybrid WC-WR with coverage guarantee 1 − α = 0.9. H LIMITATIONS H.1 SUSCEPTIBILITY TO DENSITY ESTIMATION ERRORS Given that Wasserstein regularization relies on importance-weighted conformal scores, its perfor￾mance is greatly influ…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 10 canonical work pages

  1. [1]

    Adaptive conformal prediction by reweighting nonconfor- mity score

    Salim I Amoukou and Nicolas JB Brunel. Adaptive conformal prediction by reweighting nonconfor- mity score. arXiv preprint arXiv:2303.12695,

    arXiv

  2. [5]

    We introduce a toy example to further illustrate that exchangeability does not hold

    For efficiency, all CP methods were conducted as split conformal prediction. We introduce a toy example to further illustrate that exchangeability does not hold. Consider we have two training distributions: D(1) XY = N [0, 0], 1 0 .7 0.7 1 ; D(2) XY = N [1, 1], 1 −0.6 −0.6 1 . A calibration distribution is a mixture of these two training distributions wit...

    work page 2025

  3. [6]

    0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.5 1.0 1.5 2.0 2.5 PDF: PV and Q(1) V PV Q(1) V 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 CDF: PV and Q(1) V PV Q(1) V 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.5 1.0 1.5 2.0 2.5 PDF: PV and Q(2) V PV Q(2) V 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 CDF: PV and Q(2) V PV Q(2) V Figure 6: Comparison between total variation dist...

    work page 2025

  4. [8]

    A review and comparative study on probabilistic object detection in autonomous driving

    11 Published as a conference paper at ICLR 2025 Di Feng, Ali Harakeh, Steven L Waslander, and Klaus Dietmayer. A review and comparative study on probabilistic object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(8):9961–9980,

    work page 2025

  5. [9]

    Conformal prediction with conditional guarantees

    Isaac Gibbs, John J Cherian, and Emmanuel J Cand `es. Conformal prediction with conditional guarantees. arXiv preprint arXiv:2305.12616,

    arXiv

  6. [10]

    Pierre Humbert, Batiste Le Bars, Aur´elien Bellet, and Sylvain Arlot

    URL https://arxiv.org/abs/2206.13092. Pierre Humbert, Batiste Le Bars, Aur´elien Bellet, and Sylvain Arlot. One-shot federated conformal prediction. In International Conference on Machine Learning, pp. 14153–14177. PMLR,

    arXiv

  7. [11]

    Batch multivalid conformal prediction

    Christopher Jung, Georgy Noarov, Ramya Ramalingam, and Aaron Roth. Batch multivalid conformal prediction. arXiv preprint arXiv:2209.15145,

    arXiv

  8. [12]

    An introduction to domain adaptation and transfer learning

    Wouter M Kouw and Marco Loog. An introduction to domain adaptation and transfer learning. arXiv preprint arXiv:1812.11806,

    arXiv

Show all 27 references
  1. [13]

    Multi-source conformal inference under distribution shift

    Yi Liu, Alexander W Levis, Sharon-Lise Normand, and Larry Han. Multi-source conformal inference under distribution shift. arXiv preprint arXiv:2405.09331,

    arXiv

  2. [19]

    Adaptive conformal classification with noisy labels

    Matteo Sesia, YX Wang, and Xin Tong. Adaptive conformal classification with noisy labels. arXiv preprint arXiv:2309.05092,

    arXiv

  3. [20]

    21 Published as a conference paper at ICLR 2025 0.7 0.8 0.9 1.0Coverage Airfoil 0.8 0.9 1.0 PeMSD4 0.8 0.9 1.0 PeMSD8 0.85 0.90 0.95 Seattle 0.7 0.8 0.9 1.0 US 0.8 0.9 1.0 Japan 0.00 1.00 2.00Pred. Set Size 0.00 2.00 4.00 6.00 0.00 2.00 4.00 0.00 5.00 10.00 15.00 0.00 2.50 5.0...

    2025

  4. [21]

    Provably robust conformal prediction with improved efficiency

    Ge Yan, Yaniv Romano, and Tsui-Wei Weng. Provably robust conformal prediction with improved efficiency. arXiv preprint arXiv:2404.19651,

    arXiv

  5. [22]

    We define f × g by f × g(x1, x2) = ( f (x1), g(x2)) = ( y1, y2)

    13 Published as a conference paper at ICLR 2025 A P ROOFS OF THEOREMS A.1 P ROOF OF THEOREM 1 Proof. We define f × g by f × g(x1, x2) = ( f (x1), g(x2)) = ( y1, y2). Let IdX be the identity mapping function on X , and let πi be the mapping function to the i-th marginal. The pr...

    2022

  6. [24]

    This observation proves our first reason that importance weighting effectively reduces the Wasserstein distance between calibration and test conformal scores

    Although the average coverage gaps between WR-CP and WR-CP(uw) are quite similar, at 3.1% and 2.3% respectively, the average prediction set size for WR-CP is 28.0% smaller than that of WR-CP(uw). This observation proves our first reason that importance weighting effectively re...

    2019

  7. [25]

    It includes 1503 instances

    was intentionally modified to introduce covariate shift and concept shift among them. It includes 1503 instances. The target variable is the scaled sound pressure level of NASA airfoils, and there are 5 features: log frequency, angle of attack, chord length, free-stream veloci...

    2019

  8. [1969]

    Conformal prediction is robust to label noise

    Bat-Sheva Einbinder, Stephen Bates, Anastasios N Angelopoulos, Asaf Gendler, and Yaniv Romano. Conformal prediction is robust to label noise. arXiv preprint arXiv:2209.14295, 2, 2022a. Bat-Sheva Einbinder, Yaniv Romano, Matteo Sesia, and Yanfei Zhou. Training uncertainty-aware...

    arXiv

  9. [2007]

    David Stutz, Ali Taylan Cemgil, Arnaud Doucet, et al

    URL https:// arxiv.org/abs/0706.3188. David Stutz, Ali Taylan Cemgil, Arnaud Doucet, et al. Learning optimal conformal classifiers. arXiv preprint arXiv:2110.09192,

    arXiv

  10. [2014]

    Maxime Cauchois, Suyash Gupta, and John C Duchi

    DOI: https://doi.org/10.24432/C5VW2C. Maxime Cauchois, Suyash Gupta, and John C Duchi. Knowing what you know: valid and validated confidence sets in multiclass and multilabel prediction. Journal of machine learning research, 22 (81):1–42,

    doi

  11. [2015]

    Application of uncertainty quantification to artificial intelligence in healthcare: A review of last decade (2013–2023)

    Silvia Seoni, Vicnesh Jahmunah, Massimo Salvi, Prabal Datta Barua, Filippo Molinari, and U Rajen- dra Acharya. Application of uncertainty quantification to artificial intelligence in healthcare: A review of last decade (2013–2023). Computers in Biology and Medicine, pp. 107441,

    2013

  12. [2016]

    Statistical aspects of wasserstein distances

    12 Published as a conference paper at ICLR 2025 Victor M Panaretos and Yoav Zemel. Statistical aspects of wasserstein distances. Annual review of statistics and its application, 6(1):405–431,

    2025

  13. [2018]

    Kl guided domain adaptation

    A Tuan Nguyen, Toan Tran, Yarin Gal, Philip HS Torr, and Atılım G¨unes ¸Baydin. Kl guided domain adaptation. arXiv preprint arXiv:2106.07780,

    arXiv

  14. [2019]

    Inductive confidence machines for regression

    Harris Papadopoulos, Kostas Proedrou, V olodya V ovk, and Alex Gammerman. Inductive confidence machines for regression. In Machine learning: ECML 2002: 13th European conference on machine learning Helsinki, Finland, August 19–23, 2002 proceedings 13, pp. 345–356. Springer,

    2002

  15. [2020]

    Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization

    Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731,

    1911 arXiv

  16. [2021]

    Conformal risk control

    Anastasios N Angelopoulos, Stephen Bates, Adam Fisch, Lihua Lei, and Tal Schuster. Conformal risk control. arXiv preprint arXiv:2208.02814,

    arXiv

  17. [2022]

    Distributional uncertainty propagation via optimal transport

    Liviu Aolaritei, Nicolas Lanzetti, Hongruyu Chen, and Florian D¨orfler. Distributional uncertainty propagation via optimal transport. arXiv preprint arXiv:2205.00343,

    arXiv

  18. [2023]

    A gentle introduction to conformal prediction and distribution-free uncertainty quantification

    Anastasios N Angelopoulos and Stephen Bates. A gentle introduction to conformal prediction and distribution-free uncertainty quantification. arXiv preprint arXiv:2107.07511,

    arXiv

  19. [2024]

    Normalizing flows for conformal regression

    Nicolo Colombo. Normalizing flows for conformal regression. arXiv preprint arXiv:2406.03346,

    arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.