U-aggregation: Unsupervised Aggregation of Multiple Learning Algorithms

Rui Duan

REVIEW 4 major objections 6 minor 62 references

U-aggregation: Unsupervised Aggregation of Multiple Learning Algorithms

T0 review · 4 major / 6 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 This paper claims that a label-free ensemble of pre-trained models can beat the best single model on a new population.

desk verdict A useful and mostly sound paper on unsupervised aggregation; the rank-one assumption is the load-bearing risk, and the AMP theorem doesn't cover the implemented algorithm as stated. read the letter →

arxiv 2501.18084 v1 pith:EHDZGUGV submitted 2025-01-30 stat.ML cs.LG

Rui Duan This is my paper · ORCID

classification stat.MLcs.LG MSC 60B2062H2562H3068T05

keywords ensemblelearningunsupervisedmodelaggregationheteroskedasticityrandommatrixtheoryapproximatemessagepassingspectralmethodspolygenicriskscoresgeneticprediction

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

U-aggregation is a method for combining many pre-trained prediction models when no true labels or outcomes are available in the target population. The paper's central claim is that if the models all share the same underlying risk signal after rescaling, the matrix of normalized predictions has a rank-one signal-plus-noise structure, and that structure can be recovered without supervision. The method stabilizes heteroskedastic noise at both the model and individual level using random matrix theory, then uses a sparse iterative message-passing algorithm to recover the shared risk score and per-model quality weights. In simulations it beats existing unsupervised aggregation methods when noise is heteroskedastic, and in a real polygenic risk score application the aggregated score slightly outperformed the best single model chosen with true outcomes across four traits. If correct, this gives practitioners a label-free route to robust predictions in new populations.

What carries the argument

The load-bearing mechanism is a two-stage spectral pipeline. The first stage uses the Dyson equalizer: from the diagonal of the resolvent of the symmetrized data matrix, the method estimates the rank-one row and column variance factors $h_0$ and $f$ of the noise, then bi-whitens the data so the noise becomes approximately homoskedastic. The second stage is an approximate message passing (AMP) iteration with soft-thresholding and Onsager correction terms; it alternates between estimating the sparse model-weight vector $u$ and the shared risk vector $v$. The sparsity prior on $u$ is what lets the algorithm discard non-informative or adversarial models, and the state-evolution recursion of Theorem 3.3 is what converts the iteration into a provably consistent estimator.

What would settle it

Take a collection of models whose true signals are known to be orthogonal, or simulate predictions from the model with two signal vectors $v_1$ and $v_2$ instead of one. If the normalized prediction matrix shows no clear gap between the first and second singular values and U-aggregation's consensus score correlates poorly with both signals while an oracle best model clearly wins, the central rank-one assumption fails. A simpler field check: inspect the singular-value ratios of $\bar{Y}$ before and after variance stabilization; their lack of a rank-one gap would indicate the model is not operating.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that an unsupervised consensus estimator can be consistent in a high-dimensional regime where both the number of models $d$ and sample size $n$ grow together. Under the model $Y_i = c_i(v u_i + \sigma_i F w_i)$, every informative model is a scaled copy of the common risk vector $v$ contaminated by model-specific noise; $u_i$ is a sparse indicator of informativeness. After row normalization, the data matrix decomposes as $\bar{Y} = \bar{u} v^\top + H W F$, and the element-wise variance of the noise is asymptotically rank one, $S = (1/n) h_0 f^\top$. This rank-one variance structure is what makes variance stabilization possible: Algorithm 1 estimates $h_0$ and $f$ from the resolvent of the symmetrized matrix, bi-whitens, and Algorithm 2 then applies approximate message passing with soft thresholding and Onsager corrections to recover the sparse weights and the risk vector. Theorem 3.1 gives the convergence rate of the stabilization factors, and Theorem 3.3 gives almost sure limits for the cosine similarity of the AMP estimates to the true vectors, under a phase-transition condition $\lambda^2\sqrt{\alpha} > 1$.

Load-bearing premise

The load-bearing premise is that all informative models are measuring the same underlying risk, so their predictions differ only in scale and noise; if different models capture genuinely different signals, the rank-one structure disappears and the method's estimates become biased.

Editorial extensions

If this is right

A practitioner can build a target-population risk score from public pre-trained models without collecting any outcome labels, because the learned weights track true model performance with correlations between 0.73 and 0.91 in the real application.
Larger model collections improve accuracy: the asymptotic cosine-similarity limits increase with both signal strength $\lambda$ and model-to-sample ratio $\alpha$, so adding more pre-trained models helps even when some are uninformative.
Under heteroskedastic model- and subject-level noise, U-aggregation clearly outperforms PCA and HeteroPCA, while in the homoskedastic setting it performs comparably to PCA.
The estimated weights $\hat{u}$ can serve as an unsupervised performance audit of each candidate model, which could inform documentation and ranking in public model repositories.
Below the phase-transition condition $\lambda^2\sqrt{\alpha} > 1$, consistent recovery is not guaranteed, so the method carries a detection threshold that practitioners should check.

Reading between the lines

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

Beyond the paper: the rank-one signal assumption should be tested directly on multi-signal data, such as models trained on distinct biological pathways; if the normalized prediction matrix shows several separated singular values, a rank-r extension would be needed.
Beyond the paper: the empirical result that aggregation beats the oracle best model comes from four traits in one cohort; stress-testing on traits with known population-specific genetic architecture, where models disagree in direction rather than scale, would show how far the claim generalizes.
Beyond the paper: label-free performance weights could be embedded in open model repositories as a live quality signal, letting downstream users avoid manual curation of outcomes for model selection.
Beyond the paper: because the main theorem requires Gaussian noise and a minimum signal-to-noise ratio, a practical deployment should first estimate $\lambda$ and inspect the singular-value gap, and the cross-validation procedure for $\omega$ only partially addresses model misspecification.

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.

Desk Editor's Note

read the letter

Rui Duan's paper on U-aggregation is worth your time if you care about label-free model aggregation or PRS transfer. The new piece is the combination of Dyson-equalizer variance stabilization with sparse AMP, extending the Parisi/Ahsen/Ma spectral line to heteroskedastic and adversarial settings. That is a real extension, not just a new application: it requires the random-matrix machinery they cite, and the consistency theorems move beyond prior work. The PRS application on All of Us is also genuinely useful—the method recovers model weights that correlate with true performance without seeing outcomes, and it edges out the oracle best model across four traits. That is a solid empirical result. The citation pattern is clean; it builds on the right prior work.

The soft spots are real but mostly fixable. The rank-one signal model in Equations (1)–(2) is load-bearing: every informative model is a scalar multiple of one shared v. If real models capture different latent factors, the recovered weights and consensus vector stop being interpretable as risk. The singular-value ratios in Figure 5(b) don't rule that out, since a dominant first singular value is consistent with a rank-2 signal plus noise. The Discussion acknowledges this is key, so it's not hidden.

The bigger issue is the theory-algorithm gap. Theorem 3.3 is stated for separable Lipschitz denoisers, but Algorithm 2 uses ct = ||wt||0/d, a discontinuous function of wt, plus a data-dependent threshold schedule from (16). As written, the implemented algorithm is not covered by the theorem. This is a fixable mismatch—replace the l0 count with a smooth surrogate or extend the theorem—but it means the paper's main consistency claim is not yet fully validated. Also, the proofs are deferred to a missing supplement, and assumptions (A6)–(A8) are strong and partly unverifiable in practice. None of this sinks the central idea; it makes the paper conditional.

I'd send this to a serious referee. The method is novel and the empirical application is careful; the gaps are the kind a good referee can force the author to close. If you work on aggregation or PRS, cite it—just don't lean on the full theory until the supplement appears.

Referee Report

4 major / 6 minor

Summary. The paper proposes U-aggregation, a label-free method for combining pre-trained models. It models normalized predictions as a rank-one signal plus heteroskedastic noise (Eqs. (1)-(2)), stabilizes the noise variance with a Dyson-equalizer step (Algorithm 1), and recovers sparse model weights and the consensus risk vector with an approximate message passing iteration (Algorithm 2). The main theoretical claims are consistency of the variance-stabilization estimators (Theorem 3.1) and state evolution for the AMP cosine similarities (Theorem 3.3). The method is evaluated in simulations against simple averaging, PCA, and HeteroPCA, and in a real application that aggregates PGS Catalog PRS models for height, BMI, HDL, and LDL in the All of Us cohort.

Significance. The problem is timely and practically important: publicly available pre-trained models are abundant, while target-population labels are often unavailable. The paper combines Dyson-equalizer variance stabilization with sparse AMP for unsupervised model aggregation, which is a novel combination, and the real-data PRS application is a valuable demonstration of the method's potential. Code and data availability statements are provided. If the theoretical results can be fully supplied and made to match the implemented algorithm, this would be a useful contribution to the ensemble learning and genetic risk prediction literatures.

major comments (4)

[§3.2, Theorem 3.3 and Algorithm 2] Theorem 3.3 is not a statement about the algorithm as implemented. The theorem is asserted to hold for fixed separable Lipschitz denoisers, but Algorithm 2 chooses the threshold parameters from the data through the quantile rule (16) and, when sparsity is unknown, through Algorithm 3; the state-evolution recursion (24)-(25) contains no term for this data-dependent schedule. Moreover, the Onsager coefficient c_t = ||w_t||_0 / d in Eq. (15) is not the derivative of a general separable Lipschitz denoiser, so the remark that the result holds for any such function needs qualification. The initialization text after Eq. (25) is also incoherent: it says the displayed values are 'to be substituted in Eq. (25) to yield µ0, σ0', but Eq. (25) defines (bar µ_t, bar σ_t) from (µ_t, σ_t), so the recursion lacks a well-specified starting point. Please either adapt the algorithm to the theorem or prove the theorem for the actual data-dependent procedure.
[Equations (1)-(2) and §5.1] The rank-one signal assumption is load-bearing and is not sufficiently validated. The model assumes that after scaling all informative models share a single signal vector v, so the normalized prediction matrix has the form \bar{Y} = \bar{u} v^T + HWF. In real PRS applications, models trained on different cohorts, SNP sets, and LD panels may capture partly independent genetic factors; a two-factor model \bar{Y} = \bar{u}_1 v_1^T + \bar{u}_2 v_2^T + HWF would make the AMP output \hat{v} a mixture and \hat{u}_i an unreliable measure of performance on the consensus risk. The evidence in Figure 5(b), ratios of consecutive singular values, does not rule out this scenario, since a dominant first singular value is also produced by a rank-two signal of moderate strength plus heteroskedastic noise. The Discussion correctly identifies rank-one structure as a key assumption, but the paper should provide a falsifiable diagnostic, for example testing whether the residual after removing the first singular component carries trait-relevant signal, or should explicitly characterize what U-aggregation estimates when the signal is not exactly rank one.
[§3.1 and §3.2, Theorems 3.1 and 3.3] The proofs of the two main theorems are deferred to a supplement that is not included: the text refers to 'Section ?? in the Supplement', and the simulation and real-data sections refer to Figure S1, Figure S2, and Supplementary Table 1, none of which are provided. Since Theorems 3.1 and 3.3 are the central theoretical contributions, the current version cannot be verified. The authors should include the complete supplement or move the full proofs into an appendix.
[§3.2, Assumptions (A7)-(A8)] Assumptions (A7)-(A8) are conditions on the unobserved stabilized components \tilde{u} and \tilde{v} defined in Eqs. (21)-(23), which depend on the unknown model noise levels \{\sigma_i\}, the sample noise levels \{f_j\}, and the signal strength \lambda. The paper gives no finite-sample guidance on when these conditions hold or how to check them, and the promised discussion in the Supplement is missing because of the missing supplement. In addition, Theorem 3.3 is stated under Assumptions (A1)-(A7) while the surrounding text says (A6)-(A8) are needed; please clarify which assumptions are actually required and provide concrete sufficient conditions.

minor comments (6)

[§4] The text says 'evaluated by the Person's correlation' and later 'performing even poorly than PCA'; these should be 'Pearson correlation' and 'even more poorly than PCA'.
[§4] After the sentence ending 'performing comparable with PC under homoskedastic settings' there is a stray 'A.' that should be removed.
[Algorithms 1 and 2] Algorithm 1 outputs only \tilde{Y}, but Algorithm 2 takes \hat{H} and \hat{F} as inputs from Algorithm 1; please return those diagonal matrices explicitly and state their definitions in the algorithm box.
[Equations (10)-(11)] The expressions (1/\hat{g}^{(1)} - \bar{\theta}) and (1/\hat{g}^{(2)} - \bar{\theta}) are vector operations; define them entrywise to avoid ambiguity.
[Algorithm 3] Algorithm 3 should specify exactly how \bar{Y}^{(k)} is normalized within each fold. If normalization is recomputed on each fold, the scale of the loss l(k)(ω) may differ across folds, which affects the comparability of ω values in the cross-validation criterion.
[§4] The simulation comparison would be strengthened by including at least one of the existing unsupervised aggregation baselines discussed in the Introduction, such as Ahsen et al. (2019) or Ma et al. (2023), since the paper claims improved performance over these methods in heteroskedastic settings.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: U-aggregation's derivation is model-based and evaluated against external traits; no prediction reduces to its inputs by construction.

full rationale

The paper's derivation chain is: (i) model predicted values as scaled signal plus heteroskedastic noise (Eq. 1), so normalized predictions have rank-one signal plus noise (Eq. 2); (ii) stabilize variance via Dyson-equation estimates (Algorithm 1, Theorem 3.1); (iii) recover u and v by AMP (Algorithm 2, Theorem 3.3); (iv) apply to PRS with no outcome used for fitting. Each theoretical claim is conditional on explicitly stated assumptions (A1)-(A8) and is proved using external random-matrix-theory and AMP state-evolution results (Erdos et al., Landa and Kluger, Montanari and Venkataramanan), not by assuming the conclusion. The rank-one signal assumption is substantive and load-bearing, but it is an assumption about the data-generating process, not a definitional equivalence; the Discussion openly flags it ('A key assumption underlying the method is the rank-one structure of the signal') and Section 5.1 provides singular-value evidence. Failure of this assumption would be model misspecification, not a circular reduction. No parameter is fit to target outcomes: the U-aggregation weights are obtained from normalized predictions only, and the held-out trait values are used solely for evaluation, so the 'slightly outperforms the best model' claim is not a fitted input renamed as prediction. The only self-citation in the paper (Gu, Han, and Duan 2022) appears in a survey list of transfer-learning methods in the Introduction and is not used to justify any step of the derivation. A minor implementation-theory gap (Algorithm 2 uses non-Lipschitz thresholding while Theorem 3.3 is stated for Lipschitz denoisers) is a correctness concern, not a circularity. Accordingly, no circular step can be exhibited.

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

The central claim rests on a specific statistical data model, a set of asymptotic assumptions, and the Dyson equation and AMP state evolution results. No new physical or conceptual entities are introduced.

free parameters (2)

omega (sparsity ratio s/d) = selected by cross-validation over grid (0.1,...,0.9)
Controls the soft-thresholding sequence in Algorithm 2; when unknown, Algorithm 3 picks omega by minimizing reconstruction error on the same data.
L (number of AMP iterations) = not specified
Algorithm 2 requires a stopping iteration; the paper does not give guidance. The theory requires t -> infinity and n -> infinity, so L must be chosen in practice.

assumptions (6)

domain assumption Data model (1): Y_i = c_i (u_i v + sigma_i F w_i) for each pre-trained model, with shared signal v and Gaussian noise.
The entire method is derived from this model; Section 2.1.
domain assumption After L2 normalization, the noise variance matrix of E is asymptotically rank-one: S = (1/n) h0 f^T.
Key to the Dyson equalizer variance stabilization; Section 2.2, Equations (4)-(5).
standard math Dyson equation results: for the symmetrized matrix, the diagonal of the resolvent concentrates around the solution of (7)-(8).
Used to justify Algorithm 1; cited from Erdos et al., Ajanki et al., Landa and Kluger.
ad hoc to paper AMP state evolution (Theorem 3.3) holds for the algorithm as written.
The paper states the theorem for separable Lipschitz functions, but Algorithm 2 uses an L0 norm in the Onsager correction; the proof is deferred to the supplement.
domain assumption Technical assumptions (A1)-(A8) on aspect ratio, signal delocalization, sparsity, noise bounds, and empirical distribution convergence.
These calibrate scaling so that lambda represents the signal-to-noise ratio; Section 3.1.
ad hoc to paper Choice of eta as the median singular value of Ybar in Algorithm 1.
The Dyson equalizer method requires a spectral parameter; the paper selects the median of singular values without a formal justification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of U-aggregation: Unsupervised Aggregation of Multiple Learning Algorithms." pith.science (2026). https://pith.science/paper/EHDZGUGV

@misc{pith2026250118084,
  author       = {Pith},
  title        = {Pith review of: U-aggregation: Unsupervised Aggregation of Multiple Learning Algorithms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EHDZGUGV}},
  note         = {Machine review of arXiv:2501.18084}
}

read the original abstract

Across various domains, the growing advocacy for open science and open-source machine learning has made an increasing number of models publicly available. These models allow practitioners to integrate them into their own contexts, reducing the need for extensive data labeling, training, and calibration. However, selecting the best model for a specific target population remains challenging due to issues like limited transferability, data heterogeneity, and the difficulty of obtaining true labels or outcomes in real-world settings. In this paper, we propose an unsupervised model aggregation method, U-aggregation, designed to integrate multiple pre-trained models for enhanced and robust performance in new populations. Unlike existing supervised model aggregation or super learner approaches, U-aggregation assumes no observed labels or outcomes in the target population. Our method addresses limitations in existing unsupervised model aggregation techniques by accommodating more realistic settings, including heteroskedasticity at both the model and individual levels, and the presence of adversarial models. Drawing on insights from random matrix theory, U-aggregation incorporates a variance stabilization step and an iterative sparse signal recovery process. These steps improve the estimation of individuals' true underlying risks in the target population and evaluate the relative performance of candidate models. We provide a theoretical investigation and systematic numerical experiments to elucidate the properties of U-aggregation. We demonstrate its potential real-world application by using U-aggregation to enhance genetic risk prediction of complex traits, leveraging publicly available models from the PGS Catalog.

Figures

Figures reproduced from arXiv: 2501.18084 by the authors.

**Figure 1.** Values of µt/(λσt+1) (Left) and ¯µt/( √ αλσt+1) (Right) of the AMP algorithm over iteration numbers t. The specifications of λ, α, and ω are shown in the plot. ν ∗ u is a mixture distribution, consisting of a uniform distribution over (0, c) with probability ω, and a point mass at 0 with probability 1 − ω. We choose c such that the second moment of ν ∗ u is 1. The two plots on the same row have the same parameter se… view at source ↗

**Figure 2.** Illustration of the asymptotic limits µ ∗/(λσ∗ ) (Green) and ¯µ ∗/( √ αλσ∗ ) (Blue) as functions of λ (Upper Panel) and α (Lower Panel). In the upper panel, we choose α = 0.3 and ω = 0.3. In the lower panel, we choose λ = 2, and ω = 0.5. In all settings, ν ∗ u is a mixture distribution, consisting of a uniform distribution over (0, c) with probability ω, and a point mass at 0 with probability 1 − ω. We choose c such… view at source ↗

**Figure 3.** Performance of the compared methods in estimating [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Correlation between the assigned weights and the performance of each pre-trained [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]

**Figure 5.** Figure 5: The ratios between consecutive singular values in four datasets from the AoU [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]

**Figure 6.** Figure 6: Correlation between the predicted values of each methods and the observed trait [PITH_FULL_IMAGE:figures/full_fig_p023_6.png]

**Figure 7.** Figure 7: Wordclouds showing the frequency of PRS models being the best model over the [PITH_FULL_IMAGE:figures/full_fig_p024_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 55 canonical work pages

[1]

Ahsen, M. E., R. M. Vogel, and G. A. Stolovitzky (2019). Unsupervised evaluation and weighted aggregation of ranked classification predictions. Journal of Machine Learning Research\/ 20\/ (166), 1--40

work page 2019
[2]

Erd o s, and T

Ajanki, O., L. Erd o s, and T. Kr \"u ger (2019). Quadratic vector equations on complex upper half-plane , Volume 261. American Mathematical Society

work page 2019
[3]

all of us

All of Us Research Program Investigators (2019). The “all of us” research program. New England Journal of Medicine\/ 381\/ (7), 668--676

work page 2019
[4]

BAIK, J., G. B. AROUS, and S. P \'E CH \'E (2005). Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices. The Annals of Probability\/ 33\/ (5), 1643--1697

work page 2005
[5]

Ding, and K

Bao, Z., X. Ding, and K. Wang (2021). Singular vector and singular subspace distribution for the matrix denoising model. The Annals of Statistics\/ 49\/ (1), 370--392

work page 2021
[6]

Bayati, M. and A. Montanari (2011). The dynamics of message passing on dense graphs, with applications to compressed sensing. IEEE Transactions on Information Theory\/ 57\/ (2), 764--785

work page 2011
[7]

Breiman, L. (1996). Bagging predictors. Machine learning\/ 24 , 123--140

work page 1996
[8]

B \"u hlmann, P. (2020). Invariance, causality and robustness. Statistical Science\/ 35\/ (3), 404--426

work page 2020

Show all 62 references

[9]

Cai, T. T., H. Namkoong, and S. Yadlowsky (2023). Diagnosing model performance under distribution shift. arXiv preprint arXiv:2303.02011\/

2023 arXiv
[10]

Chen, Y. and E. J. Cand \`e s (2018). The projected power method: An efficient algorithm for joint alignment from pairwise differences. Communications on Pure and Applied Mathematics\/ 71\/ (8), 1648--1714

2018
[11]

Choi, S. W., T. S.-H. Mak, and P. F. O’Reilly (2020). Tutorial: a guide to performing polygenic risk score analyses. Nature protocols\/ 15\/ (9), 2759--2772

2020
[12]

Coombes, B. J., A. Ploner, S. E. Bergen, and J. M. Biernacka (2020). A principal component approach to improve association testing with polygenic risk scores. Genetic epidemiology\/ 44\/ (7), 676--686

2020
[13]

Dong, X., Z. Yu, W. Cao, Y. Shi, and Q. Ma (2020). A survey on ensemble learning. Frontiers of Computer Science\/ 14 , 241--258

2020
[14]

Gavish, and E

Donoho, D., M. Gavish, and E. Romanov (2023). Screenot: Exact mse-optimal singular value thresholding in correlated noise. The Annals of Statistics\/ 51\/ (1), 122--148

2023
[15]

Donoho, D. L., A. Maleki, and A. Montanari (2009). Message-passing algorithms for compressed sensing. Proceedings of the National Academy of Sciences\/ 106\/ (45), 18914--18919

2009
[16]

Duncan, L., H. Shen, B. Gelaye, J. Meijsen, K. Ressler, M. Feldman, R. Peterson, and B. Domingue (2019). Analysis of polygenic risk score usage and performance in diverse human populations. Nature communications\/ 10\/ (1), 3328

2019
[17]

u ger, and D. Schr \

Erd o s, L., T. Kr \"u ger, and D. Schr \"o der (2019). Random matrices with slow correlation decay. In Forum of Mathematics, Sigma , Volume 7, pp.\ e8. Cambridge University Press

2019
[18]

Chikowore, A

Fatumo, S., T. Chikowore, A. Choudhury, M. Ayub, A. R. Martin, and K. Kuchenbaecker (2022). A roadmap to increase diversity in genomic studies. Nature medicine\/ 28\/ (2), 243--250

2022
[19]

Leeb, and E

Gavish, M., W. Leeb, and E. Romanov (2023). Matrix denoising with partial noise statistics: optimal singular value shrinkage of spiked f-matrices. Information and Inference: A Journal of the IMA\/ 12\/ (3), 2020--2065

2023
[20]

Gabelloni, A

Ghezzi, A., D. Gabelloni, A. Martini, and A. Natalicchio (2018). Crowdsourcing: a review and suggestions for future research. International Journal of management reviews\/ 20\/ (2), 343--363

2018
[21]

Han, and R

Gu, T., Y. Han, and R. Duan (2022). Robust angle-based transfer learning in high dimensions. arXiv preprint arXiv:2210.12759\/

2022 arXiv
[22]

Pleiss, Y

Guo, C., G. Pleiss, Y. Sun, and K. Q. Weinberger (2017). On calibration of modern neural networks. In International conference on machine learning , pp.\ 1321--1330. PMLR

2017
[23]

Guo, Z. (2024). Statistical inference for maximin effects: Identifying stable associations across multiple studies. Journal of the American Statistical Association\/ 119\/ (547), 1968--1984

2024
[24]

Hector, E. C. and R. Martin (2024). Turning the information-sharing dial: efficient inference from different data sources. Electronic Journal of Statistics\/ 18\/ (2), 2974--3020

2024
[25]

Gilman, L

Hong, D., K. Gilman, L. Balzano, and J. A. Fessler (2021). Heppcat: Probabilistic pca for data with heteroscedastic noise. IEEE Transactions on Signal Processing\/ 69 , 4819--4834

2021
[26]

Huang, Y., W. Li, F. Macheret, R. A. Gabriel, and L. Ohno-Machado (2020). A tutorial on calibration measurements and calibration models for clinical prediction models. Journal of the American Medical Informatics Association\/ 27\/ (4), 621--633

2020
[27]

Johnstone, I. M. (2001). On the distribution of the largest eigenvalue in principal components analysis. The Annals of Statistics\/ 29\/ (2), 295--327

2001
[28]

Khera, A. V., M. Chaffin, K. G. Aragam, M. E. Haas, C. Roselli, S. H. Choi, P. Natarajan, E. S. Lander, S. A. Lubitz, P. T. Ellinor, et al. (2018). Genome-wide polygenic scores for common diseases identify individuals with risk equivalent to monogenic mutations. Nature genetic...

2018
[29]

Roth, and K

Klementiev, A., D. Roth, and K. Small (2008). Unsupervised rank aggregation with distance-based models. In Proceedings of the 25th international conference on Machine learning , pp.\ 472--479

2008
[30]

Kullo, I. J., C. M. Lewis, M. Inouye, A. R. Martin, S. Ripatti, and N. Chatterjee (2022). Polygenic scores in biomedical research. Nature Reviews Genetics\/ 23\/ (9), 524--532

2022
[31]

Lambert, S. A., L. Gil, S. Jupp, S. C. Ritchie, Y. Xu, A. Buniello, A. McMahon, G. Abraham, M. Chapman, H. Parkinson, et al. (2021). The polygenic score catalog as an open database for reproducibility and systematic evaluation. Nature Genetics\/ 53\/ (4), 420--425

2021
[32]

Lambert, S. A., B. Wingfield, J. T. Gibson, L. Gil, S. Ramachandran, F. Yvon, S. Saverimuttu, E. Tinsley, E. Lewis, S. C. Ritchie, et al. (2024). The polygenic score catalog: new functionality and tools to enable fair research. medRxiv\/ , 2024--05

2024
[33]

Landa, B. and Y. Kluger (2023). The dyson equalizer: Adaptive noise stabilization for low-rank signal detection and recovery. arXiv preprint arXiv:2306.11263\/

2023 arXiv
[34]

Landa, B., T. T. Zhang, and Y. Kluger (2022). Biwhitening reveals the rank of a count matrix. SIAM journal on mathematics of data science\/ 4\/ (4), 1420--1446

2022
[35]

Lecu \'e , G. and P. Rigollet (2014). Optimal learning with q-aggregation

2014
[36]

Leeb, W. and E. Romanov (2021). Optimal spectral shrinkage and pca with heteroscedastic noise. IEEE Transactions on Information Theory\/ 67\/ (5), 3009--3037

2021
[37]

Li, S., T. T. Cai, and H. Li (2022). Transfer learning for high-dimensional linear regression: Prediction, estimation and minimax optimality. Journal of the Royal Statistical Society Series B: Statistical Methodology\/ 84\/ (1), 149--173

2022
[38]

Lobach, D. F. and D. E. Detmer (2007). Research challenges for electronic health records. American Journal of Preventive Medicine\/ 32\/ (5), S104--S111

2007
[39]

Ma, R., E. D. Sun, and J. Zou (2023). A spectral method for assessing and combining multiple data visualizations. Nature Communications\/ 14\/ (1), 780

2023
[40]

Martin, A. R., C. R. Gignoux, R. K. Walters, G. L. Wojcik, B. M. Neale, S. Gravel, M. J. Daly, C. D. Bustamante, and E. E. Kenny (2017). Human demographic history impacts genetic risk prediction across diverse populations. The American Journal of Human Genetics\/ 100\/ (4), 635--649

2017
[41]

McKiernan, E. C., P. E. Bourne, C. T. Brown, S. Buck, A. Kenall, J. Lin, D. McDougall, B. A. Nosek, K. Ram, C. K. Soderberg, et al. (2016). How open science helps researchers succeed. elife\/ 5 , e16800

2016
[42]

Meinshausen, N. and P. Bühlmann (2015). Maximin effects in inhomogeneous large-scale data. The Annals of Statistics\/ 43\/ (4), 1801--1830

2015
[43]

Parisi, and M

M \'e zard, M., G. Parisi, and M. A. Virasoro (1987). Spin glass theory and beyond: An Introduction to the Replica Method and Its Applications , Volume 9. World Scientific Publishing Company

1987
[44]

Montanari, A. and R. Venkataramanan (2021). Estimation of low-rank matrices via approximate message passing. The Annals of Statistics\/ 49\/ (1)

2021
[45]

Onaran, E. and S. Villar (2017). Projected power iteration for network alignment. In Wavelets and Sparsity XVII , Volume 10394, pp.\ 298--305. SPIE

2017
[46]

Strino, B

Parisi, F., F. Strino, B. Nadler, and Y. Kluger (2014). Ranking and combining multiple predictors without labeled data. Proceedings of the National Academy of Sciences\/ 111\/ (4), 1253--1258

2014
[47]

Rahimian, H. and S. Mehrotra (2019). Distributionally robust optimization: A review. arXiv preprint arXiv:1908.05659\/

2019 arXiv
[48]

Richardson, T. and R. Urbanke (2008). Modern coding theory . Cambridge university press

2008
[49]

Schapire, R. E. (2003). The boosting approach to machine learning: An overview. Nonlinear estimation and classification\/ , 149--171

2003
[50]

Sonnenburg, S., M. L. Braun, C. S. Ong, S. Bengio, L. Bottou, G. Holmes, Y. LeCunn, K.-R. Muller, F. Pereira, C. E. Rasmussen, et al. (2007). The need for open source software in machine learning

2007
[51]

Tsybakov, A. B. (2004). Optimal aggregation of classifiers in statistical learning. The Annals of Statistics\/ 32\/ (1), 135--166

2004
[52]

Van der Laan, M. J., E. C. Polley, and A. E. Hubbard (2007). Super learner. Statistical applications in genetics and molecular biology\/ 6\/ (1)

2007
[53]

Vaughan, J. W. (2018). Making better use of the crowd: How crowdsourcing can advance machine learning research. Journal of Machine Learning Research\/ 18\/ (193), 1--46

2018
[54]

Volkovs, M. N. and R. S. Zemel (2014). New learning methods for supervised and unsupervised preference aggregation. The Journal of Machine Learning Research\/ 15\/ (1), 1135--1176

2014
[55]

Zhong, and Z

Wang, T., X. Zhong, and Z. Fan (2024). Universality of approximate message passing algorithms and tensor networks. The Annals of Applied Probability\/ 34\/ (4), 3943--3994

2024
[56]

Weiss, K., T. M. Khoshgoftaar, and D. Wang (2016). A survey of transfer learning. Journal of Big data\/ 3 , 1--40

2016
[57]

Wojcik, G. L., M. Graff, K. K. Nishimura, R. Tao, J. Haessler, C. R. Gignoux, H. M. Highland, Y. M. Patel, E. P. Sorokin, C. L. Avery, et al. (2019). Genetic analyses of diverse populations improves discovery for complex traits. Nature\/ 570\/ (7762), 514--518

2019
[58]

Wray, N. R., M. E. Goddard, and P. M. Visscher (2007). Prediction of individual genetic risk to disease from genome-wide association studies. Genome research\/ 17\/ (10), 1520--1528

2007
[59]

Guo, and T

Xiong, X., Z. Guo, and T. Cai (2023). Distributionally robust transfer learning. arXiv preprint arXiv:2309.06534\/

2023 arXiv
[60]

Xu, Q., Y. Yuan, J. Wang, and A. Qu (2024). Crowdsourcing utilizing subgroup structure of latent factor modeling. Journal of the American Statistical Association\/ 119\/ (546), 1192--1204

2024
[61]

Shahbaba, N

Yuan, Y., B. Shahbaba, N. Fortin, K. Cooper, Q. Nie, and A. Qu (2024). Optimal transport for latent integration with an application to heterogeneous neuronal activity data. arXiv preprint arXiv:2407.00099\/

2024 arXiv
[62]

Zhang, A. R., T. T. Cai, and Y. Wu (2022). Heteroskedastic pca: Algorithm, optimality, and applications. The Annals of Statistics\/ 50\/ (1), 53--80

2022

Pith tools

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

[1] [1]

Ahsen, M. E., R. M. Vogel, and G. A. Stolovitzky (2019). Unsupervised evaluation and weighted aggregation of ranked classification predictions. Journal of Machine Learning Research\/ 20\/ (166), 1--40

work page 2019

[2] [2]

Erd o s, and T

Ajanki, O., L. Erd o s, and T. Kr \"u ger (2019). Quadratic vector equations on complex upper half-plane , Volume 261. American Mathematical Society

work page 2019

[3] [3]

all of us

All of Us Research Program Investigators (2019). The “all of us” research program. New England Journal of Medicine\/ 381\/ (7), 668--676

work page 2019

[4] [4]

BAIK, J., G. B. AROUS, and S. P \'E CH \'E (2005). Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices. The Annals of Probability\/ 33\/ (5), 1643--1697

work page 2005

[5] [5]

Ding, and K

Bao, Z., X. Ding, and K. Wang (2021). Singular vector and singular subspace distribution for the matrix denoising model. The Annals of Statistics\/ 49\/ (1), 370--392

work page 2021

[6] [6]

Bayati, M. and A. Montanari (2011). The dynamics of message passing on dense graphs, with applications to compressed sensing. IEEE Transactions on Information Theory\/ 57\/ (2), 764--785

work page 2011

[7] [7]

Breiman, L. (1996). Bagging predictors. Machine learning\/ 24 , 123--140

work page 1996

[8] [8]

B \"u hlmann, P. (2020). Invariance, causality and robustness. Statistical Science\/ 35\/ (3), 404--426

work page 2020