Pith. sign in

REVIEW 3 major objections 5 minor 55 references

Conformal Changepoint Localization and Root Cause Analysis with Corrupted Observations

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 Corrupted observations can be downweighted in conformal changepoint and root-cause localization, preserving finite-sample coverage while shrinking confidence sets by up to an order of magnitude under heavy contamination.

desk verdict Coverage guarantees are sound, but the efficiency gains rest on an unproven uncertainty proxy; worth refereeing but needs revision. read the letter →

arxiv 2607.26481 v1 pith:6DLPUNW5 submitted 2026-07-29 cs.LG eess.SP

Seunghun Yu , Meiyi Zhu , Petar Popovski , Joonhyuk Kang , Osvaldo Simeone This is my paper
classification cs.LGeess.SP
keywords changepointlocalizationrootcauseanalysisconformalpredictionHubercontaminationuncertaintyquantificationmeta-learningdistribution-freeinferenceconfidencesets
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 asks whether conformal changepoint localization and root-cause analysis can remain both statistically valid and informative when observations are corrupted. It establishes that the split-exchangeability assumption underlying the unweighted CONCH and CROC procedures survives a Huber-style contamination model, so their coverage guarantees are preserved—but the reported confidence sets can become impractically large. The proposed weighted variants, W-CONCH and W-CROC, multiply each observation's likelihood contribution by a weight derived from a lower bound on the contaminated marginal density; the ideal weight is the posterior probability that the observation is clean, and in practice it is read off classifier uncertainty signals such as evidential deep learning or Monte Carlo dropout. A meta-learned extension optimizes a differentiable surrogate of the confidence-set size, removing the need to know the contamination level in advance. On image and telecom benchmarks, weighting reduces average confidence-set size by up to an order of magnitude at 70% contamination while empirical coverage stays near the nominal level.

What carries the argument

The load-bearing object is the weighted changepoint-plausibility (CPP) score S_t^w(X) = Σ_{i≤t} w_0(X_i) Δ_i − Σ_{i≤ξ̂_t(X)} w_(t)(X_i) Δ_i, where Δ_i = log( p̂(0|X_i)/p̂(1|X_i) ) is the classifier log-posterior ratio and w_j(X_i) ∈ [0,1] downweights suspicious observations. The weights are ideally the posterior probability that X_i is clean, derived from a formal ELBO lower bound on the contaminated mixture density; in practice they are computed from classifier uncertainty signals using side-specific quantile thresholds. The key property carrying the argument is permutation-equivariance of the weighted score under within-segment splits, which preserves the finite-sample rank argument underl

What would settle it

Construct a changepoint task where corrupted observations are adversarial examples with high classifier confidence and clean samples near the class boundary have high uncertainty. Run W-CONCH with soft weighting at ε=0.7: if the average confidence-set size stays close to unweighted CONCH while the oracle weighting still shrinks it, the uncertainty-as-corruption proxy fails; conversely, if the meta-learned variant still shrinks the set, the method learns a better signal than raw uncertainty.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the split-exchangeability property required by conformal changepoint localization survives Huber contamination, so coverage remains valid even when a large fraction of observations is corrupted; the remaining problem is purely one of informativeness. To restore informativeness, the paper derives a weighted changepoint-plausibility score from a lower bound on the contaminated marginal density, in which the ideal weight is the posterior probability that an observation is clean. Propositions 2 and 4 state that the weighted procedures W-CONCH and W-CROC retain the marginal coverage guarantees Pr(ξ ∈ C) ≥ 1−α and Pr(d⋆ ∈ K) ≥ 1−α for any choice of the weighti

Load-bearing premise

The load-bearing premise is that classifier uncertainty reliably marks corrupted observations: clean samples are confident, corrupted samples are uncertain. If high-confidence adversarial corruptions or inherently uncertain clean samples violate this, the weights discard informative observations and the confidence-set-size gains disappear.

Editorial extensions

If this is right

  • Under contamination, unweighted CONCH and CROC sets remain valid but grow quickly; weighting restores informativeness without changing the marginal coverage guarantee.
  • The coverage guarantee holds for any hyperparameters λ and β, so tuning those parameters affects only efficiency, not validity.
  • Meta-learned weights trained across a mixture of contamination levels can operate without knowing the test-time contamination level, approaching oracle performance on DomainNet changepoint and CIFAR-100 root-cause tasks.
  • Prior information about where the changepoint or root cause is likely to lie can be folded in through level allocation, with the coverage bound weakening gracefully from 1−α to 1−α_max.

Reading between the lines

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

  • The efficiency gains rest on an unproven proxy: classifier uncertainty is assumed to separate clean from corrupted observations. I would test whether high-confidence adversarial corruptions, or clean samples near the decision boundary, break the weighting and leave sets as large as unweighted CONCH.
  • Because the only requirement for coverage is permutation-equivariance of the score, the same weighted-score construction could transfer to other conformal set predictors whose scores decompose over exchangeable segments, such as online or multi-stream settings.
  • The weights are tuned to minimize average set size, so an operator should treat the resulting set as marginally, not conditionally, valid; per-sequence guarantees would require a different construction.
  • A quantitative statement about how set size degrades with contamination level—rather than only experiments—would let a practitioner decide when the unweighted procedure is already good enough.
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 offline changepoint localization and root-cause analysis when observations are corrupted under a Huber-type contamination model. It first observes that split exchangeability, and hence the finite-sample coverage guarantees of CONCH and CROC, is preserved under such contamination (Lemma 1, Lemmas 2 and Proposition 4). It then proposes W-CONCH and W-CROC, which use classifier-based uncertainty signals to downweight observations that are likely to be corrupted, with the aim of reducing the size of the conformal confidence set. The paper also introduces MW-CONCH and MW-CROC, meta-learning variants that optimize a differentiable surrogate of the confidence-set size to learn the uncertainty-to-weight mapping. Experiments on DomainNet, CIFAR-100, MNIST, and the Milan telecom dataset report substantial average confidence-set size reductions at high contamination levels while maintaining nominal coverage.

Significance. The coverage-preservation result under contamination (Lemma 1) is clean and useful, and the extension to weighted scores is a natural and practically motivated idea. If the efficiency claims hold, the paper provides a meaningful step toward keeping conformal changepoint/root-cause methods informative under realistic data corruption. The meta-learning framework is well structured and the paper includes detailed algorithmic pseudocode (Algorithm 1) and extensive experiments. However, the theoretical derivation of the weighted score contains a load-bearing mismatch (Proposition 1 versus the implemented score), and the central efficiency claim rests on an unverified proxy relating classifier uncertainty to contamination probability. These issues need to be addressed before the paper's main contribution can be considered established.

major comments (3)
  1. [§IV-A2, Proposition 1 and Definition 1] Proposition 1 states g_j(x) ≥ C·f_j(x)w_j(x), but the proof in Appendix A-B derives log g_j(x) ≥ w_j(x) log f_j(x) − C_R, i.e., g_j(x) ≥ C'·f_j(x)^{w_j(x)}. The implemented weighted CPP score in (14) and (18) uses multiplicative weighting f_j(x)w_j(x), not the power weighting f_j(x)^{w_j(x)}. Thus the stated bound does not justify the score actually used. Moreover, in the proof w_j(x) is defined as r_j(Y=0|x) for an arbitrary auxiliary posterior r_j, not the true clean posterior; the claim after (14) that w_j ideally corresponds to the posterior probability that the sample is clean is not established. The authors must correct this: either state the proposition correctly and derive the score from it, or explicitly present multiplicative weighting as a heuristic approximation.
  2. [§IV-B, §VI] The entire efficiency gain of W-CONCH/W-CROC rests on the unproven assumption that classifier uncertainty h(X_i) is a reliable proxy for the probability that an observation is corrupted. Section IV-B simply asserts that a classifier 'should ideally be confident on clean in-distribution samples and uncertain on contaminated ones.' The experiments corrupt observations only with Gaussian blur/noise on the chosen benchmarks, which broadly follows this heuristic. Under high-confidence adversarial perturbations, or when clean samples near decision boundaries have high uncertainty, the proposed weights can downweight informative clean observations, potentially leaving the confidence set as large as unweighted CONCH. The paper should either prove a condition under which the uncertainty ranking is monotone in the contamination posterior, or add stress tests with non-Gaussian corruptions and repor
  3. [§VI-B1 vs Appendix B-A] The text states that MW-CONCH (EDL-F) at ε=0.7 on DomainNet reaches size 3.61, 'even outperforming oracle-based W-CONCH' (9.81), and interprets this as meta-learning learning 'classifier representations that are more robust to contamination.' However, Appendix B-A says that during meta-learning the backbone is frozen and only the selection head is updated, so the classifier log-posterior ratio Δ_i used in the score (18) is not changed by meta-learning. This contradicts the interpretation. It is also surprising that a method without access to true contamination indicators outperforms an oracle that uses them. The authors should clarify the experimental setup (e.g., whether the oracle uses the same classifier and whether weights are applied consistently) and either provide a validated explanation or temper the claim.
minor comments (5)
  1. [Appendix A-B] In the proof of Proposition 1, the text says 'Substituting (39) and (A-B) into (38)' but the second equation should be numbered (40). Please fix the cross-reference.
  2. [Table I] The table footnote says that for ε=0 the corresponding CONCH result is reported because β=0 is undefined, but the W-CONCH (Hard) and W-CONCH (Soft) entries at ε=0 differ from the CONCH entry (e.g., DomainNet: 1.51 vs 2.07). This is inconsistent; please clarify or correct the reported numbers.
  3. [Proposition 1] The proposition assumes a bounded observation space and q bounded away from zero and infinity. These assumptions are never discussed in the experiments, where the classifiers operate on images and RBF features. Please comment on whether these assumptions hold approximately in the tested settings, or note that the proposition is only a motivating oracle result.
  4. [Algorithm 1] In line 12, the notation uses ε_tr^(ℓ) to index the training-level loss, while the surrounding text writes ε. Please unify the notation to avoid confusion.
  5. [Figures 5 and 6] The abbreviations EDL-F, EDL-M, MC-F, MC-M are used in the figure legends but fully defined only in the caption text. Please define them explicitly in the main text or in a legend below each figure.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; coverage proofs are standard conformal rank arguments, and the efficiency gains rest on an unverified but non-circular uncertainty heuristic.

full rationale

The central validity claims are self-contained. Lemma 1 derives split exchangeability of the Huber-contaminated sequence by applying the deterministic contamination map to the augmented exchangeable sequence; Lemma 2 and Propositions 2/4 then invoke the standard finite-sample rank bound for permutation-equivariant scores. The weighted score is permutation-equivariant because the side-wise thresholds in (22) are empirical quantiles of uncertainty scores and are invariant to within-segment permutations. No prediction in the paper is identical by construction to a fitted input: the hard/soft weights use uncertainty scores from pretrained or meta-learned models, and meta-learning optimizes a smoothed surrogate of set size on training tasks and is evaluated on disjoint held-out tasks. The strongest caveat is that the efficiency improvement assumes classifier uncertainty is a reliable proxy for contamination (Section IV-B); if violated, the method may not reduce set size. That is a correctness/robustness limitation, not circular reasoning. Self-citations ([27], [30], [38], [43], [50]) appear in related work or as implementation techniques and are not load-bearing for the coverage theorems; the core CONCH/CROC results are external prior work [13], [14].

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central coverage theorem rests on standard exchangeability and independence of contamination. The efficiency claims rest on the unproved uncertainty-proxy heuristic and on hyperparameters matched to the test contamination level in the non-meta variants.

free parameters (5)
  • β (quantile hyperparameter) = Set to ε in hard/soft experiments (0.1, 0.3, 0.5, 0.7); 0.3 or 0.1 for mixed-εtr meta-learning
    Controls the quantile threshold separating clean from corrupted observations. The headline non-meta results require matching β to the true test contamination level ε.
  • λ (soft-weighting temperature) = 0.05 (DomainNet, Milan), 0.01 (CIFAR-100 W-CONCH), 0.005 (CIFAR MW-CROC MC dropout)
    Sigmoid sharpness for soft weights; tuned per dataset and per method.
  • Smoothing temperatures τ_q, τ_1, τ_2 (meta-learning) = τ_q=1.0; τ1,τ2 init 0.5 annealed ×0.95/epoch, min 0.01 (DomainNet) or 0.05 (CIFAR)
    Control the differentiable surrogates of quantiles, indicators, and set-size counting during meta-training.
  • EDL RBF head scales γ, σ (Milan) = Initialized γ=e^2, σ=1 and learned in log space
    Radial-basis evidence model used for the telecom dataset; additional fitted parameters beyond the core method.
  • Prior weight v and likely-set size k (prior-informed CROC) = v from 1 to 5, k=1..3 in the experiments
    Strength and cardinality of the user-provided prior over root-cause streams.
assumptions (6)
  • domain assumption Split exchangeability of the clean sequence at the true changepoint (Assumption 1)
    Inherited from CONCH [13]; without it the finite-sample coverage proof collapses. Stated in Section II-A.
  • domain assumption i.i.d. contamination indicators Y_i independent of clean data, and i.i.d. contamination draws Z_i (Huber model)
    Necessary for Lemma 1. If contamination is adversarial or dependent on the clean signal, split-exchangeability need not be preserved.
  • domain assumption Bounded observation space and |log q(x)| ≤ C_q for the contamination density (Proposition 1)
    Used to derive the lower bound (13); not strictly needed for coverage but needed for the stated theoretical motivation of weighting.
  • domain assumption A pre-trained classifier on clean labeled data, independent of the evaluation batch, whose log-posterior ratio approximates the likelihood ratio
    Invoked in Section IV-A3 to replace densities with classifier scores; efficiency depends on this approximation.
  • ad hoc to paper Classifier uncertainty is a reliable proxy for the probability that an observation is clean
    The central efficiency heuristic (Section IV-B). No proof is given; empirically checked only on the chosen contamination types.
  • ad hoc to paper The multiplicative weighting f_j(x)w_j(x) in the score (14) follows from the bound in Proposition 1
    The proof of Proposition 1 yields a power form f_j(x)^{w_j(x)}, not f_j(x)·w_j(x); the paper does not reconcile this mismatch.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conformal Changepoint Localization and Root Cause Analysis with Corrupted Observations." pith.science (2026). https://pith.science/paper/6DLPUNW5

@misc{pith2026260726481,
  author       = {Pith},
  title        = {Pith review of: Conformal Changepoint Localization and Root Cause Analysis with Corrupted Observations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6DLPUNW5}},
  note         = {Machine review of arXiv:2607.26481}
}
read the original abstract

Detecting when the statistical behavior of an engineered system changes, and identifying which component is responsible, are core problems in the monitoring of telecommunication networks, robotic platforms, security infrastructure, and multi-agent systems. In safety- and mission-critical deployments, such decisions must be accompanied by statistical reliability guarantees rather than by point estimates alone. Conformal changepoint localization (CONCH) and conformal root cause analysis (CROC) meet this need by returning confidence sets that contain the true changepoint, or the true root-cause stream, with a user-specified probability, without parametric assumptions on the data-generating process. In practice, however, observations are frequently corrupted, e.g., by outliers, sensor faults, or adversarial perturbations. While the finite-sample coverage of these procedures is preserved under contamination, the resulting confidence sets can become uninformatively large. Adopting a Huber-type contamination model, this paper proposes weighted CONCH (W-CONCH) and weighted CROC (W-CROC), which downweight observations that are likely to be corrupted with the goal of reducing confidence set size when data may be corrupted. The weighting mechanism, derived from a formal bound on the unknown corrupted data densities, leverages pre-existing second-order classifier-based uncertainty signals, such as those produced by evidential deep learning or Bayesian learning. W-CONCH and W-CROC are further generalized by introducing a meta-learning procedure for the weights that optimizes a differentiable surrogate of the confidence set size. Experiments on image-based and real-world changepoint and root-cause benchmarks show that uncertainty-based weighting substantially reduces confidence set size while maintaining the target coverage.

Figures

Figures reproduced from arXiv: 2607.26481 by the authors.

Figure 1
Figure 1. Illustration of the offline changepoint localizatio [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the problem of offline root-cause loc [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Operation of MW-CONCH: The uncertainty model [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: W-CONCH on DomainNet (left panels) and CIFAR-100 (ri [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Performance against the contamination levels [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: W-CROC and prior-informed CROC on CIFAR-100 with [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 6 linked inside Pith

  1. [1]

    Selective review of offline change point detection methods,

    C. Truong, L. Oudre, and N. V ayatis, “Selective review of offline change point detection methods,” Signal Process. , vol. 167, p. 107299, 2020

    2020

  2. [2]

    A survey of methods for time series change point detection,

    S. Aminikhanghahi and D. J. Cook, “A survey of methods for time series change point detection,” Knowl. Inf. Syst. , vol. 51, no. 2, pp. 339–367, 2017

    2017

  3. [3]

    Survey on models and techniques for root-cause analysis,

    M. Sol´ e, V . Munt´ es-Mulero, A. I. Rana, and G. Estrada, “ Survey on models and techniques for root-cause analysis,” arXiv preprint arXiv:1701.08546, 2017

    arXiv 2017

  4. [4]

    Anomaly detection and failure r oot cause analysis in (micro)service-based cloud applicat ions: A survey,

    J. Soldani and A. Brogi, “Anomaly detection and failure r oot cause analysis in (micro)service-based cloud applicat ions: A survey,” ACM Comput. Surv., vol. 55, no. 3, pp. 1–39, 2022

    2022

  5. [5]

    Detection and localizati on of change-points in high-dimensional network traffic dat a,

    C. L´ evy-Leduc and F. Roueff, “Detection and localizati on of change-points in high-dimensional network traffic dat a,” Ann. Appl. Stat. , vol. 3, no. 2, pp. 637–662, 2009

    2009

  6. [6]

    Understanding O-RAN: Architecture, interfaces, algori thms, security, and research challenges,

    M. Polese, L. Bonati, S. D’oro, S. Basagni, and T. Melodia , “Understanding O-RAN: Architecture, interfaces, algori thms, security, and research challenges,” IEEE Commun. Surv. Tutor ., vol. 25, no. 2, pp. 1376–1411, 2023. 33

    2023

  7. [7]

    Conformal calibrat ion: Ensuring the reliability of black-box AI in wireless sy stems,

    O. Simeone, S. Park, and M. Zecchin, “Conformal calibrat ion: Ensuring the reliability of black-box AI in wireless sy stems,” IEEE Commun. Mag., 2026

    2026

  8. [8]

    A novel approach to detection of intrusions in computer networks via adaptive sequential and batch-sequential change-point de tection methods,

    A. G. Tartakovsky, B. L. Rozovskii, R. B. Blaˇ zek, and H. K im, “A novel approach to detection of intrusions in computer networks via adaptive sequential and batch-sequential change-point de tection methods,” IEEE Trans. Signal Process. , vol. 54, no. 9, pp. 3372–3382, 2006

    2006

Show all 55 references
  1. [9]

    On fault detection and diag nosis in robotic systems,

    E. Khalastchi and M. Kalech, “On fault detection and diag nosis in robotic systems,” ACM Comput. Surv. , vol. 51, no. 1, pp. 1–24, 2018

    2018

  2. [10]

    Nonparametric d etection of anomalous data streams,

    S. Zou, Y . Liang, H. V . Poor, and X. Shi, “Nonparametric d etection of anomalous data streams,” IEEE Trans. Signal Process. , vol. 65, no. 21, pp. 5785–5797, 2017

    2017

  3. [11]

    Asymptotically optimal anomaly d etection via sequential testing,

    K. Cohen and Q. Zhao, “Asymptotically optimal anomaly d etection via sequential testing,” IEEE Trans. Signal Process. , vol. 63, no. 11, pp. 2929–2941, 2015

    2015

  4. [12]

    Decision theoretic foundations for conformal prediction: Optimal u ncertainty quantification for risk-averse agents,

    S. Kiyani, G. Pappas, A. Roth, and H. Hassani, “Decision theoretic foundations for conformal prediction: Optimal u ncertainty quantification for risk-averse agents,” arXiv preprint arXiv:2502.02561 , 2025

    2025 arXiv

  5. [13]

    Conformal changepoint localiza tion,

    R. Hore and A. Ramdas, “Conformal changepoint localiza tion,” arXiv preprint arXiv:2602.06267 , 2026

    2026 arXiv

  6. [14]

    Distribution-free root cause analysis,

    ——, “Distribution-free root cause analysis,” arXiv preprint arXiv:2605.21627 , 2026

    2026 arXiv

  7. [15]

    A test for a change in a parameter occurring a t an unknown point,

    E. S. Page, “A test for a change in a parameter occurring a t an unknown point,” Biometrika, vol. 42, no. 3/4, pp. 523–527, 1955

    1955

  8. [16]

    The likelihood ratio test fo r a change-point in simple linear regression,

    H.-J. Kim and D. Siegmund, “The likelihood ratio test fo r a change-point in simple linear regression,” Biometrika, vol. 76, no. 3, pp. 409–423, 1989

    1989

  9. [17]

    A non-parametric approach to the change -point problem,

    A. N. Pettitt, “A non-parametric approach to the change -point problem,” J. R. Stat. Soc. Ser . C (Appl. Stat.) , vol. 28, no. 2, pp. 126–135, 1979

    1979

  10. [18]

    Practical and powerful kernel-bas ed change-point detection,

    H. Song and H. Chen, “Practical and powerful kernel-bas ed change-point detection,” IEEE Trans. Signal Process. , vol. 72, pp. 5174–5186, 2024

    2024

  11. [19]

    A log-linear nonparametric online changepoint detection algorithm bas ed on functional pruning,

    G. Romano, I. A. Eckley, and P . Fearnhead, “A log-linear nonparametric online changepoint detection algorithm bas ed on functional pruning,” IEEE Trans. Signal Process. , vol. 72, pp. 594–606, 2023

    2023

  12. [20]

    Adversarially robust change point dete ction,

    M. Li and Y . Y u, “Adversarially robust change point dete ction,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , 2021, pp. 22 955–22 967

    2021

  13. [21]

    Online change poi nt detection under heavy-tailedness and contamination,

    E. Y . N. Tang, Y . Chen, M. Li, and Y . Y u, “Online change poi nt detection under heavy-tailedness and contamination,” arXiv preprint arXiv:2606.09737, 2026

    2026 arXiv

  14. [22]

    Root cause analysis for microservice systems via hierarchical reinforcement learning from huma n feedback,

    L. Wang, C. Zhang, R. Ding, Y . Xu, Q. Chen, W. Zou, Q. Chen, M. Zhang, X. Gao, H. Fan et al. , “Root cause analysis for microservice systems via hierarchical reinforcement learning from huma n feedback,” in Proc. 29th ACM SIGKDD Conf. Knowl. Discovery Data Mining (KDD), 2023,...

    2023

  15. [23]

    Machine-learn ing applications of algorithmic randomness,

    V . V ovk, A. Gammerman, and C. Saunders, “Machine-learn ing applications of algorithmic randomness,” in Proc. 16th Int. Conf. Mach. Learn. (ICML) . Morgan Kaufmann, 1999, pp. 444–453

    1999

  16. [24]

    A tutorial on conformal predicti on,

    G. Shafer and V . V ovk, “A tutorial on conformal predicti on,” J. Mach. Learn. Res. , vol. 9, no. 3, 2008

    2008

  17. [25]

    Conformalized quantile regression,

    Y . Romano, E. Patterson, and E. Candes, “Conformalized quantile regression,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , 2019

    2019

  18. [26]

    Conformal prediction under covariate shift,

    R. J. Tibshirani, R. F. Barber, E. J. Cand` es, and A. Ramd as, “Conformal prediction under covariate shift,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , 2019

    2019

  19. [27]

    Ca librating wireless ai via meta-learned context-dependent conformal prediction,

    S. Y oo, S. Park, P . Popovski, J. Kang, and O. Simeone, “Ca librating wireless ai via meta-learned context-dependent conformal prediction,” IEEE Trans. Signal Process. , 2026

    2026

  20. [28]

    Conformal prediction beyond exchangeability,

    R. F. Barber, E. J. Cand` es, A. Ramdas, and R. J. Tibshira ni, “Conformal prediction beyond exchangeability,” Ann. Statist. , vol. 51, no. 2, pp. 816–845, 2023

    2023

  21. [29]

    Adaptive conformal inferen ce under distribution shift,

    I. Gibbs and E. J. Cand` es, “Adaptive conformal inferen ce under distribution shift,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , 2021

    2021

  22. [30]

    Localized adaptive risk con trol,

    M. Zecchin and O. Simeone, “Localized adaptive risk con trol,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , 2024, pp. 8165–8192

    2024

  23. [31]

    Testing randomness online,

    V . V ovk, “Testing randomness online,” Statist. Sci. , vol. 36, no. 4, pp. 595–611, 2021

    2021

  24. [32]

    Robust estimation of a location parameter ,

    P . J. Huber, “Robust estimation of a location parameter ,” Ann. Math. Statist. , vol. 35, no. 1, pp. 73–101, 1964

    1964

  25. [33]

    F. R. Hampel, E. M. Ronchetti, P . J. Rousseeuw, and W. A. S tahel, Robust Statistics: The Approach Based on Influence Function s. New Y ork: John Wiley & Sons, 1986

    1986

  26. [34]

    Certified de fenses for data poisoning attacks,

    J. Steinhardt, P . W. W. Koh, and P . S. Liang, “Certified de fenses for data poisoning attacks,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017. 34

    2017

  27. [35]

    Evidential deep learning to quantify classification uncertainty,

    M. Sensoy, L. Kaplan, and M. Kandemir, “Evidential deep learning to quantify classification uncertainty,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , 2018

    2018

  28. [36]

    Dropout as a bayesian approxi mation: Representing model uncertainty in deep learning,

    Y . Gal and Z. Ghahramani, “Dropout as a bayesian approxi mation: Representing model uncertainty in deep learning,” in Proc. Int. Conf. Mach. Learn. (ICML) . PMLR, 2016, pp. 1050–1059

    2016

  29. [37]

    Sim ple and scalable predictive uncertainty estimation using d eep ensembles,

    B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Sim ple and scalable predictive uncertainty estimation using d eep ensembles,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , 2017

    2017

  30. [38]

    Calibrating Bayesia n learning via regularization, confidence minimization, an d selective inference,

    J. Huang, S. Park, and O. Simeone, “Calibrating Bayesia n learning via regularization, confidence minimization, an d selective inference,” IEEE Trans. Signal Process. , vol. 73, pp. 4492–4505, 2025

    2025

  31. [39]

    Simeone, Machine learning for engineers

    O. Simeone, Machine learning for engineers . Cambridge Univ. Press, 2022

    2022

  32. [40]

    Model-agnostic meta -learning for fast adaptation of deep networks,

    C. Finn, P . Abbeel, and S. Levine, “Model-agnostic meta -learning for fast adaptation of deep networks,” in Proc. Int. Conf. Mach. Learn. (ICML). PMLR, 2017, pp. 1126–1135

    2017

  33. [41]

    Learning with limited samples: Meta-learning an d applications to communication systems,

    L. Chen, S. T. Jose, I. Nikoloska, S. Park, T. Chen, and O. Simeone, “Learning with limited samples: Meta-learning an d applications to communication systems,” F ound. Trends Signal Process., vol. 17, no. 2, pp. 79–208, 2023

    2023

  34. [42]

    Learning optimal conformal classifiers,

    D. Stutz, A. T. Cemgil, A. Doucet et al. , “Learning optimal conformal classifiers,” arXiv preprint arXiv:2110.09192 , 2021

    2021 arXiv

  35. [43]

    Few-shot calibrat ion of set predictors via meta-learned cross-validation-b ased conformal prediction,

    S. Park, K. M. Cohen, and O. Simeone, “Few-shot calibrat ion of set predictors via meta-learned cross-validation-b ased conformal prediction,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 46, no. 1, pp. 280–291, 2023

    2023

  36. [44]

    J. A. Rice, Mathematical statistics and data analysis . Thomson/Brooks/Cole Belmont, CA, 2007, vol. 371

    2007

  37. [45]

    Offline changepoint lo calization using a matrix of conformal p-values,

    S. Dandapanthula and A. Ramdas, “Offline changepoint lo calization using a matrix of conformal p-values,” arXiv preprint arXiv:2505.00292, 2025

    2025

  38. [46]

    A. D. Joseph, B. Nelson, B. I. Rubinstein, and J. Tygar, Adversarial machine learning . Cambridge University Press, 2018

    2018

  39. [47]

    Predictive uncertainty estim ation via prior networks,

    A. Malinin and M. Gales, “Predictive uncertainty estim ation via prior networks,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , 2018

    2018

  40. [48]

    False disc overy control with p-value weighting,

    C. R. Genovese, K. Roeder, and L. Wasserman, “False disc overy control with p-value weighting,” Biometrika, vol. 93, no. 3, pp. 509–524, 2006

    2006

  41. [49]

    A multi-source dataset of urban life in the city of Milan and th e province of Trentino,

    G. Barlacchi, M. De Nadai, R. Larcher, A. Casella, C. Chi tic, G. Torrisi, F. Antonelli, A. V espignani, A. Pentland, a nd B. Lepri, “A multi-source dataset of urban life in the city of Milan and th e province of Trentino,” Sci. Data , vol. 2, no. 1, p. 150055, 2015

    2015

  42. [50]

    Gen eralization and informativeness of conformal prediction,

    M. Zecchin, S. Park, O. Simeone, and F. Hellstr¨ om, “Gen eralization and informativeness of conformal prediction, ” in Proc. IEEE Int. Symp. Inf. Theory (ISIT) . IEEE, 2024, pp. 244–249

    2024

  43. [51]

    Moment matching for multi-source domain adaptation,

    X. Peng, Q. Bai, X. Xia, Z. Huang, K. Saenko, and B. Wang, “ Moment matching for multi-source domain adaptation,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , 2019, pp. 1406–1415

    2019

  44. [52]

    Learning multiple layers of features f rom tiny images,

    A. Krizhevsky, “Learning multiple layers of features f rom tiny images,” University of Toronto, Tech. Rep., 2009

    2009

  45. [53]

    Deep residual learni ng for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learni ng for image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 770–778

    2016

  46. [54]

    Uncer tainty estimation using a single deep deterministic neural network,

    J. V an Amersfoort, L. Smith, Y . W. Teh, and Y . Gal, “Uncer tainty estimation using a single deep deterministic neural network,” in Proc. Int. Conf. Mach. Learn. (ICML) . PMLR, 2020, pp. 9690–9700

    2020

  47. [55]

    The mnist database of handwritten digit image s for machine learning research [best of the web],

    L. Deng, “The mnist database of handwritten digit image s for machine learning research [best of the web],” IEEE Signal Process. Mag. , vol. 29, no. 6, pp. 141–142, 2012

    2012

Pith tools

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