Pith. sign in

REVIEW 5 minor 51 references

Group Distributionally Robust Optimization with Flexible Sample Queries

T0 review · 0 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper proves that group distributionally robust optimization can be solved with an arbitrary, time-varying number of samples per round, matching the sample complexity of the fixed 1-or-m sample regimes.

desk verdict A solid, genuinely useful generalization of GDRO to flexible sample sizes with correct high-probability bounds; the novelty claim needs a sharper comparison and the experiments need error bars. read the letter →

arxiv 2505.15212 v1 pith:4ZROV4AL submitted 2025-05-21 cs.LG math.OC

Haomin Bai , Dingzhi Yu , Shuai Li , Haipeng Luo , Lijun Zhang This is my paper · ORCID
classification cs.LGmath.OC
keywords groupdistributionallyrobustoptimizationpredictionwithlimitedadviceonlineconvexfollow-the-regularized-leaderhigh-probabilityregrettime-varyingsamplesizesanytimealgorithmcomplexity
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

Group distributionally robust optimization (GDRO) seeks a model that performs well across m distributions at once. This paper establishes that the usual rigid sampling schemes—exactly 1 or exactly m samples per round—are unnecessary: a two-player online-learning algorithm can accept any number r_t between 1 and m per round, varying over time, and still achieve optimization error O(1/t $\sqrt$(sum_{j=1}^t (m/r_j) log m)) with high probability. If correct, this means practitioners can adapt sampling to fluctuating budgets (bandwidth, annotation cost, group availability) without sacrificing the known O(m log m/$epsilon^{2}$) sample complexity for any fixed r. The paper also gives the first high-probability regret bound for non-oblivious prediction with limited advice, a by-product of independent interest.

What carries the argument

The load-bearing object is the unified prediction-with-limited-advice (PLA) strategy for the q-player: a single FTRL update on the simplex with step size eta_{q,t} = sqrt(ln m / (m sum_{j<=t} 1/r_j)), a biased IX estimator when r_t = 1, an unbiased estimator when r_t >= 2, and dependent rounding to select the r_t - 1 additional groups. This unifies Exp3-IX (for single-sample rounds) and PrLiA (for multi-sample rounds) into one process, and its analysis produces the first high-probability regret bound for non-oblivious PLA. The w-player uses FTRL with the ghost-iterate technique to handle non-oblivious stochastic gradients, and the two regret bounds are tied together by a decomposition that bounds the optimization error by Regret_w/t + Regret'_q/t plus a martingale concentration term.

What would settle it

Take a convex problem with m groups, fix a sample size r (say r=5), run Algorithm 3 for T rounds, and measure the worst-group excess risk max_{i in [m]} R_i(w_bar_T) minus the true optimum across many repeated trials; if the empirical tail probability of exceeding C * sqrt(m log m / (r T)) for a fixed constant C stays far above delta, or if the gap does not shrink at the 1/sqrt(T) rate, the central claim is contradicted.

Watch

Extended reading notes

Core claim

The central claim is that GDRO with flexible sample queries reduces to a two-player game in which the q-player faces non-oblivious prediction with limited advice (PLA), and that this PLA problem admits a high-probability regret bound of O($\sqrt$(sum_{j=1}^t (m/r_j) log m)) against an adaptive adversary. The paper constructs a unified strategy that runs a single follow-the-regularized-leader (FTRL) update on the simplex with step size eta_{q,t} = $\sqrt$(ln m / (m sum_{j<=t} 1/r_j)), using an implicit-exploration (IX) loss estimator when r_t = 1 and an unbiased estimator when r_t >= 2, with the extra indices drawn by dependent rounding. Feeding this into the two-player game and pairing it with an FTRL-based w-player whose regret is controlled by a ghost-iterate argument yields Theorem 1: with probability at least 1-delta, the optimization error is O(1/t $\sqrt$(sum_{j=1}^t (m/r_j) log m)). For fixed r this becomes O($\sqrt$(m log m / (r t))), giving sample complexity O(m log m / $epsilon^{2}$) for every r in [m], matching the r=1 and r=m regimes.

Load-bearing premise

Everything rests on Assumption 4: each group risk R_i(w) is convex in w, so the original problem equals the convex-concave minimax problem (2) and the regret-based decomposition in Lemma 1 holds; if the losses are non-convex, the main error bound no longer follows.

Editorial extensions

If this is right

  • GDRO sampling can track fluctuating resource budgets: if a round only affords r_t of the m groups, the algorithm uses r_t, and the error bound degrades gracefully as the term 1/r_t in the square-root sum.
  • For any fixed sample size r, the algorithm matches the previous best sample complexity O(m log m / epsilon^2), interpolating between the 1-sample and m-sample regimes.
  • The FTRL-with-averaged-output design removes the O(log t) factor present in SMD-based anytime GDRO, improving the fixed-r error bound to O(sqrt(m log m / (r t))) without knowing the horizon in advance.
  • The unified PLA analysis gives a high-probability regret bound for non-oblivious PLA, closing a gap left by prior expected-regret and oblivious-only analyses.

Reading between the lines

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

  • Because the q-player is treated as a black box, the same unified PLA algorithm could be dropped into other minimax stochastic optimization problems that need flexible expert or group queries, not just GDRO.
  • The O(log log t) cost in the time-uniform guarantee (Theorem 2) suggests that anytime deployment—stopping at an arbitrary round—incurs almost no penalty, making the method natural for streaming and federated settings.
  • The convexity assumption (Assumption 4) is the real gate: if group risk functions are non-convex, the two-player equivalence and the regret decomposition break, so extending the approach to deep models would require convex surrogates or a stationarity-based analysis.
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

0 major / 5 minor

Summary. The paper studies group distributionally robust optimization (GDRO) when the number of samples queried per round, r_t, can vary arbitrarily in [m]. It casts the problem as a two-player game: the w-player solves a non-oblivious online convex optimization problem with stochastic gradients, and the q-player solves a non-oblivious prediction-with-limited-advice (PLA) problem. The authors propose a unified FTRL-based PLA algorithm, together with a hybrid Exp3-IX/PrLiA variant, and prove high-probability regret bounds of the form O(sqrt(sum_{j=1}^t (m/r_j) log m)). Combining this with an FTRL strategy for the w-player, they obtain an optimization error bound of O((1/t) sqrt(sum_{j=1}^t (m/r_j) log m)) (Theorem 1), which for fixed r gives O(sqrt(m log m / (r t))) and sample complexity O(m log m / epsilon^2). A time-uniform version is also proved for oblivious r_t sequences (Theorem 2). Experiments on a synthetic binary dataset and a real-world diabetes dataset compare the proposed algorithms with existing fixed-sample-size methods.

Significance. The paper fills a clear gap: previous GDRO algorithms were restricted to one sample or m samples per round, while this work supports arbitrary and time-varying sample sizes. The flexible-query bound generalizes the prior r=1 and r=m results and, through the use of FTRL with uniformly averaged outputs, removes an O(log t) factor relative to earlier anytime SMD algorithms. The proofs are detailed and follow standard martingale, FTRL, and ghost-iterate techniques; no free parameters are fitted to data, and the main statements are backed by explicit concentration arguments. The convexity assumption (Assumption 4) is explicit and standard, though it limits applicability to non-convex models; the experiments are correspondingly restricted to linear logistic loss. The absence of released code and of error bars weakens the empirical contribution but does not affect the theoretical claims. Overall, if the bounds are correct, this is a solid and useful theoretical contribution to GDRO and to online learning with limited advice.

minor comments (5)
  1. [Remark 3 / Eq. (23)] The O(1/t sqrt(sum m/r_j log m)) rate is stated without an explicit asymptotic regime. The displayed bound (23) contains terms such as (m/t) sqrt(ln m) (sum m/r_j)^{1/4} sqrt(ln(20/delta)) and (1/t) m^2 / sqrt(ln m) ln(20/delta), which are not dominated by the leading term until t is sufficiently large. Please state explicitly that the O-notation is for fixed m and delta as t tends to infinity, or give a quantitative threshold for when the leading term dominates.
  2. [Abstract / Section 3.3] The paper claims the 'first high-probability regret bound for non-oblivious PLA,' but Section 2.2 itself cites Yun et al. (2018) as exploring PLA in the non-oblivious setting. Please clarify precisely what is new: the varying sample-size model, the unified FTRL estimator, or the high-probability guarantee in this specific feedback model. This would also help readers assess the novelty claim.
  3. [Appendix C.6, term(d) / Eq. (47)] In bounding term(d), the text says 'setting alpha_{j,i}=2 gamma_j in Lemma 15' and then divides by 2 gamma_t. To make the division immediate, one should set alpha_{j,i}=2 gamma_t, which is allowed because gamma is non-increasing, before applying Lemma 15. Please correct the notation so the argument matches the displayed inequality.
  4. [Remark 7 / Appendix A] Remark 7 states that a time-uniform version satisfies Regret'_{q,t} = O((1/t) sqrt(sum m/r_j max{...})). Since Regret'_{q,t} is not normalized by t, the displayed order for the regret should be O(sqrt(sum m/r_j ...)); the (1/t) factor belongs to the resulting optimization error bound. Please fix this inconsistency.
  5. [Section 4] The experiments are repeated only 5 times and no standard deviations, confidence intervals, or code release are provided. The qualitative conclusions comparing HYB/UNI with Online(1) and Online(1)' would be substantially easier to assess with error bars, and a code release would support reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central bound follows from explicit step sizes, external concentration inequalities, and re-derived regret decompositions.

full rationale

The derivation chain is self-contained. Theorem 1 substitutes Lemma 2 and Lemma 3 into Lemma 1. Lemma 1 (Appendix C.5) re-derives the regret decomposition from Jensen's inequality and Azuma/Bernstein concentration; it does not assume the desired optimization error. Lemma 2 (Appendix C.6) bounds Regret'_q,t from the FTRL guarantee (37), the loss-estimator decompositions (39)-(40), and martingale concentration (Lemmas 7, 8, 15, 16); the supporting lemmas are proved in Appendix C using external concentration inequalities and Seldin et al.'s combinatorial lemmas, with no fitted constants. Lemma 3 (Appendix C.7) uses the FTRL regret formula and the ghost-iterate construction of Nemirovski et al. (2009), re-proved for FTRL. Step sizes are explicit: eta_q,t = sqrt(ln m / (m * sum_j 1/r_j)) and eta_w,t = sqrt(2)D/(sqrt(5)G) * 1/sqrt(t), so no parameter is fitted to data whose error is later reported. Citations to Zhang et al. (2023) are used as baselines and motivation; the decomposition that citation is said to support is independently proved in Lemma 1. Assumption 4 (convexity) is an explicit scope condition used in Jensen's inequality, not an input that forces the rate. No equation reduces to its own conclusion, and no load-bearing premise is imported solely from a self-citation. The minor self-citations are comparative and non-load-bearing.

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

The central claim rests on the four standard GDRO assumptions (convexity, bounded gradients, bounded losses, bounded domain) and on the availability of an oracle that reveals r_t at the start of each round. No free parameters are fitted and no new entities are introduced.

assumptions (5)
  • domain assumption Assumption 1: domain W is convex and its nu_w-diameter is bounded by D.
    Used in Lemma 3 and the FTRL analysis to bound the regularizer term.
  • domain assumption Assumption 2: the gradient norm is bounded by G for all sample draws.
    Used to bound the stochastic gradients and the ghost-iterate martingale differences.
  • domain assumption Assumption 3: loss ell(w;z) is in [0,1] for all w and z.
    Defines the transformed losses \hat s_t in (7) and controls the variance of the loss estimators.
  • domain assumption Assumption 4: each risk function R_i(w) is convex.
    Needed for the minimax formulation and for the Jensen-based regret decomposition in Lemma 1.
  • domain assumption For the time-uniform guarantee (Theorem 2), the sequence {r_t} is oblivious, i.e., fixed in advance.
    The time-uniform concentration argument requires the sample-size sequence to be independent of the algorithm's decisions; the fixed-t bound (Theorem 1) does not need this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Group Distributionally Robust Optimization with Flexible Sample Queries." pith.science (2026). https://pith.science/paper/4ZROV4AL

@misc{pith2026250515212,
  author       = {Pith},
  title        = {Pith review of: Group Distributionally Robust Optimization with Flexible Sample Queries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4ZROV4AL}},
  note         = {Machine review of arXiv:2505.15212}
}
abstract

Group distributionally robust optimization (GDRO) aims to develop models that perform well across $m$ distributions simultaneously. Existing GDRO algorithms can only process a fixed number of samples per iteration, either 1 or $m$, and therefore can not support scenarios where the sample size varies dynamically. To address this limitation, we investigate GDRO with flexible sample queries and cast it as a two-player game: one player solves an online convex optimization problem, while the other tackles a prediction with limited advice (PLA) problem. Within such a game, we propose a novel PLA algorithm, constructing appropriate loss estimators for cases where the sample size is either 1 or not, and updating the decision using follow-the-regularized-leader. Then, we establish the first high-probability regret bound for non-oblivious PLA. Building upon the above approach, we develop a GDRO algorithm that allows an arbitrary and varying sample size per round, achieving a high-probability optimization error bound of $O\left(\frac{1}{t}\sqrt{\sum_{j=1}^t \frac{m}{r_j}\log m}\right)$, where $r_t$ denotes the sample size at round $t$. This result demonstrates that the optimization error decreases as the number of samples increases and implies a consistent sample complexity of $O(m\log (m)/\epsilon^2)$ for any fixed sample size $r\in[m]$, aligning with existing bounds for cases of $r=1$ or $m$. We validate our approach on synthetic binary and real-world multi-class datasets.

Figures

Figures reproduced from arXiv: 2505.15212 by the authors.

Figure 1
Figure 1. Illustration of GDRO with fixed sample sizes ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. GDRO under varying sample sizes: max risk versus the number of iterations [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. GDRO under varying sample sizes: max risk versus the running time [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: HYB under different fixed sample sizes: max risk versus the number of iterations [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: HYB under different fixed sample sizes: max risk versus the number of samples [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: UNI under different fixed sample sizes: max risk versus the number of iterations [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: UNI under different fixed sample sizes: max risk versus the number of samples [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 46 canonical work pages

  1. [1]

    Minimax regret optimization for robust machine learning under distribution shift

    Alekh Agarwal and Tong Zhang. Minimax regret optimization for robust machine learning under distribution shift. In Proceedings of 35th Conference on Learning Theory, pages 2704--2729, 2022

    work page 2022

  2. [2]

    Budgeted prediction with expert advice

    Kareem Amin, Satyen Kale, Gerald Tesauro, and Deepak Turaga. Budgeted prediction with expert advice. Proceedings of the AAAI Conference on Artificial Intelligence, 29 0 (1), 2015

    work page 2015

  3. [3]

    Schapire

    Peter Auer, Nicol\` o Cesa-Bianchi, Yoav Freund, and Robert E. Schapire. The nonstochastic multiarmed bandit problem. SIAM Journal on Computing, 32 0 (1): 0 48--77, 2002

    work page 2002

  4. [4]

    Regret analysis of stochastic and nonstochastic multi-armed bandit problems

    S\' e bastien Bubeck and Nicol\` o Cesa-Bianchi. Regret analysis of stochastic and nonstochastic multi-armed bandit problems. Foundations and Trends in Machine Learning, 5 0 (1): 0 1--122, 2012

    work page 2012

  5. [5]

    Distributionally robust optimization via ball oracle acceleration

    Yair Carmon and Danielle Hausler. Distributionally robust optimization via ball oracle acceleration. In Advances in Neural Information Processing Systems 35, pages 35866--35879, 2022

    work page 2022

  6. [6]

    Cesa-Bianchi and G

    N. Cesa-Bianchi and G. Lugosi. Prediction, Learning, and Games. Cambridge University Press, 2006

    work page 2006

  7. [7]

    Anytime online-to-batch, optimism and acceleration

    Ashok Cutkosky. Anytime online-to-batch, optimism and acceleration. In Proceedings of the 36th International Conference on Machine Learning, pages 1446--1454, 2019

    work page 2019

  8. [8]

    Distributionally robust optimization under moment uncertainty with application to data-driven problems

    Erick Delage and Yinyu Ye. Distributionally robust optimization under moment uncertainty with application to data-driven problems. Operations research, 58 0 (3): 0 595--612, 2010

    2010

Show all 51 references
  1. [9]

    Duchi and Hongseok Namkoong

    John C. Duchi and Hongseok Namkoong. Learning models with uniform performance via distributionally robust optimization. The Annals of Statistics, 49 0 (3): 0 1378--1406, 2021

    2021

  2. [10]

    Probability: theory and examples, volume 49

    Rick Durrett. Probability: theory and examples, volume 49. Cambridge University Press, 2019

    2019

  3. [11]

    Freedman

    David A. Freedman. On tail probabilities for martingales. The Annals of Probability, 3 0 (1): 0 100--118, 1975

    1975

  4. [12]

    A survey on concept drift adaptation

    Joao Gama, Indre Zliobaite, Albert Bifet, Mykol Pechenizkiy, and Abdelhamid Bouchachia. A survey on concept drift adaptation. ACM Computing Surveys, 46 0 (4): 0 1--37, 2014

    2014

  5. [13]

    Dependent rounding and its applications to approximation algorithms

    Rajiv Gandhi, Samir Khuller, Srinivasan Parthasarathy, and Aravind Srinivasan. Dependent rounding and its applications to approximation algorithms. Journal of the ACM, 53 0 (3): 0 324--360, 2006

    2006

  6. [14]

    Statistical machine learning: A unified framework

    Richard Golden. Statistical machine learning: A unified framework. Chapman and Hall/CRC, 2020

    2020

  7. [15]

    Jordan, and Eric Zhao

    Nika Haghtalab, Michael I. Jordan, and Eric Zhao. On-demand sampling: Learning optimally from multiple distributions. In Advances in Neural Information Processing Systems 35, pages 406--419, 2022

    2022

  8. [16]

    Multiarmed bandits with limited expert advice

    Satyen Kale. Multiarmed bandits with limited expert advice. In Proceedings of The 27th Conference on Learning Theory, pages 107--122, 2014

    2014

  9. [17]

    Efficient learning by implicit exploration in bandit problems with side observations

    Tom \' a s Koc \' a k, Gergely Neu, Michal Valko, and R \' e mi Munos. Efficient learning by implicit exploration in bandit problems with side observations. In Advances in Neural Information Processing Systems 27, pages 613--621, 2014

    2014

  10. [18]

    Generative models improve fairness of medical classifiers under distribution shifts

    Ira Ktena, Olivia Wiles, Isabela Albuquerque, Sylvestre-Alvise Rebuffi, Ryutaro Tanno, Abhijit Guha Roy, Shekoofeh Azizi, Danielle Belgrave, Pushmeet Kohli, Taylan Cemgil, Alan Karthikesalingam, and Sven Gowal. Generative models improve fairness of medical classifiers under di...

    2024

  11. [19]

    Stable prediction across unknown environments

    Kun Kuang, Peng Cui, Susan Athey, Ruoxuan Xiong, and Bo Li. Stable prediction across unknown environments. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1617--1626, 2018

    2018

  12. [20]

    Minimax optimal reinforcement learning with quasi-optimism

    Harin Lee and Min hwan Oh. Minimax optimal reinforcement learning with quasi-optimism. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  13. [21]

    Lasso bandit with compatibility condition on optimal arm

    Harin Lee, Taehyun Hwang, and Min hwan Oh. Lasso bandit with compatibility condition on optimal arm. ArXiv e-prints, arXiv:2406.00823, 2025

    2025 arXiv

  14. [22]

    Agnostic federated learning

    Mehryar Mohri, Gary Sivek, and Ananda Theertha Suresh. Agnostic federated learning. In Proceedings of the 36th International Conference on Machine Learning, pages 4615--4625, 2019

    2019

  15. [23]

    Stochastic gradient methods for distributionally robust optimization with f-divergences

    Hongseok Namkoong and John C Duchi. Stochastic gradient methods for distributionally robust optimization with f-divergences. In Advances in Neural Information Processing Systems 29, pages 2208--2216, 2016

    2016

  16. [24]

    Nemirovski, A

    A. Nemirovski, A. Juditsky, G. Lan, and A. Shapiro. Robust stochastic approximation approach to stochastic programming. SIAM Journal on Optimization, 19 0 (4): 0 1574--1609, 2009

    2009

  17. [25]

    Explore no more: Improved high-probability regret bounds for non-stochastic bandits

    Gergely Neu. Explore no more: Improved high-probability regret bounds for non-stochastic bandits. In Advances in Neural Information Processing Systems 28, pages 3168--3176, 2015

    2015

  18. [26]

    Mehta, and Cristóbal Guzmán

    Quan Nguyen, Nishant A. Mehta, and Cristóbal Guzmán. Beyond minimax rates in group distributionally robust optimization via a novel notion of sparsity. ArXiv e-prints, arXiv:2410.00690, 2024

    2024 arXiv

  19. [27]

    A modern introduction to online learning

    Francesco Orabona. A modern introduction to online learning. ArXiv e-prints, arXiv:1912.13213, 2023

    1912 arXiv

  20. [28]

    To cool or not to cool? temperature network meets large foundation models via DRO

    Zi-Hao Qiu, Siqi Guo, Mao Xu, Tuo Zhao, Lijun Zhang, and Tianbao Yang. To cool or not to cool? temperature network meets large foundation models via DRO . In Proceedings of the 41st International Conference on Machine Learning, pages 41604--41643, 2024

    2024

  21. [29]

    Frameworks and results in distributionally robust optimization

    Hamed Rahimian and Sanjay Mehrotra. Frameworks and results in distributionally robust optimization. Open Journal of Mathematical Optimization, 3: 0 1--85, 2022

    2022

  22. [30]

    Group robust preference optimization in reward-free rlhf

    Shyam Sundhar Ramesh, Yifan Hu, Iason Chaimalas, Viraj Mehta, Pier Giuseppe Sessa, Haitham Bou Ammar, and Ilija Bogunovic. Group robust preference optimization in reward-free rlhf. ArXiv e-prints, arXiv:2405.20304, 2024

    2024 arXiv

  23. [31]

    Efficient online-bandit strategies for minimax learning problems

    Christophe Roux, Elias Wirth, Sebastian Pokutta, and Thomas Kerdreux. Efficient online-bandit strategies for minimax learning problems. ArXiv e-prints, arXiv:2105.13939, 2021

    2021 arXiv

  24. [32]

    Fast rates for prediction with limited expert advice

    El Mehdi Saad and Gilles Blanchard. Fast rates for prediction with limited expert advice. In Advances in Neural Information Processing Systems 34, pages 23582--23591, 2021

    2021

  25. [33]

    Constant regret for sequence prediction with limited advice

    El Mehdi Saad and Gilles Blanchard. Constant regret for sequence prediction with limited advice. In Proceedings of The 34th International Conference on Algorithmic Learning Theory, pages 1343--1386, 2023

    2023

  26. [34]

    Hashimoto, and Percy Liang

    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. In International Conference on Learning Representations, 2020

    2020

  27. [35]

    Prediction with limited advice and multiarmed bandits with paid observations

    Yevgeny Seldin, Peter Bartlett, Koby Crammer, and Yasin Abbasi-Yadkori. Prediction with limited advice and multiarmed bandits with paid observations. In Proceedings of the 31st International Conference on Machine Learning, pages 280--287, 2014

    2014

  28. [36]

    On distributionally robust optimization and data rebalancing

    Agnieszka S owik and L\' e on Bottou. On distributionally robust optimization and data rebalancing. In Proceedings of the 25th International Conference on Artificial Intelligence and Statistics, pages 1283--1297, 2022

    2022

  29. [37]

    Optimal algorithms for group distributionally robust optimization and beyond

    Tasuku Soma, Khashayar Gatmiry, and Stefanie Jegelka. Optimal algorithms for group distributionally robust optimization and beyond. ArXiv e-prints, arXiv:2212.13669, 2022

    2022 arXiv

  30. [38]

    DeShazo, Chris Gennings, Juan L

    Beata Strack, Jonathan P. DeShazo, Chris Gennings, Juan L. Olmo, Sebastian Ventura, Krzysztof J. Cios, and John N. Clore. Impact of hba1c measurement on hospital readmission rates: Analysis of 70,000 clinical database patient records. BioMed Research International, 2014 0 (1):...

    2014

  31. [39]

    Covariate shift adaptation by importance weighted cross validation

    Masashi Sugiyama, Matthias Krauledat, and Klaus-Robert M \"u ller. Covariate shift adaptation by importance weighted cross validation. Journal of Machine Learning Research, 8 0 (35): 0 985--1005, 2007

    2007

  32. [40]

    Adaptation to easy data in prediction with limited advice

    Tobias Sommer Thune and Yevgeny Seldin. Adaptation to easy data in prediction with limited advice. In Advances in Neural Information Processing Systems 31, pages 2909--2918, 2018

    2018

  33. [41]

    Algorithms for adversarial bandit problems with multiple plays

    Taishi Uchiya, Atsuyoshi Nakamura, and Mineichi Kudo. Algorithms for adversarial bandit problems with multiple plays. In Algorithmic Learning Theory, pages 375--389, 2010

    2010

  34. [42]

    Vladimir N. Vapnik. The Nature of Statistical Learning Theory. Springer, second edition, 2000

    2000

  35. [43]

    Kozat Kozat

    Nuri Mert Vural, Hakan Gokcesu, Kaan Gokcesu, and Suleyman S. Kozat Kozat. Minimax optimal algorithms for adversarial bandit problem with multiple plays. IEEE Transactions on Signal Processing, 67 0 (16): 0 4383--4398, 2019

    2019

  36. [44]

    Towards unbiased information extraction and adaptation in cross-domain recommendation

    Yibo Wang, Yingchun Jian, Wenhao Yang, Shiyin Lu, Lei Shen, Bing Wang, Xiaoyi Zeng, and Lijun Zhang. Towards unbiased information extraction and adaptation in cross-domain recommendation. Proceedings of the AAAI Conference on Artificial Intelligence, 39 0 (12): 0 12757--12765, 2025

    2025

  37. [45]

    Doremi: Optimizing data mixtures speeds up language model pretraining

    Sang Michael Xie, Hieu Pham, Xuanyi Dong, Nan Du, Hanxiao Liu, Yifeng Lu, Percy S Liang, Quoc V Le, Tengyu Ma, and Adams Wei Yu. Doremi: Optimizing data mixtures speeds up language model pretraining. In Advances in Neural Information Processing Systems 36, pages 69798--69818, 2023

    2023

  38. [46]

    Efficient algorithms for empirical group distributionally robust optimization and beyond

    Dingzhi Yu, Yunuo Cai, Wei Jiang, and Lijun Zhang. Efficient algorithms for empirical group distributionally robust optimization and beyond. In Proceedings of the 41st International Conference on Machine Learning, pages 57384--57414, 2024

    2024

  39. [47]

    Multi-armed bandit with additional observations

    Donggyu Yun, Alexandre Proutiere, Sumyeong Ahn, Jinwoo Shin, and Yung Yi. Multi-armed bandit with additional observations. Proceedings of the ACM on Measurement and Analysis of Computing Systems, 2 0 (1): 0 1--22, 2018

    2018

  40. [48]

    Stochastic approximation approaches to group distributionally robust optimization

    Lijun Zhang, Peng Zhao, Zhenhua Zhuang, Tianbao Yang, and Zhi-Hua Zhou. Stochastic approximation approaches to group distributionally robust optimization. In Advances in Neural Information Processing Systems 36, pages 52490--52522, 2023

    2023

  41. [49]

    Efficient stochastic approximation of minimax excess risk optimization

    Lijun Zhang, Haomin Bai, Wei-Wei Tu, Ping Yang, and Yao Hu. Efficient stochastic approximation of minimax excess risk optimization. In Proceedings of the 41st International Conference on Machine Learning, pages 58599--58630, 2024 a

    2024

  42. [50]

    Stochastic approximation approaches to group distributionally robust optimization and beyond

    Lijun Zhang, Haomin Bai, Peng Zhao, Tianbao Yang, and Zhi-Hua Zhou. Stochastic approximation approaches to group distributionally robust optimization and beyond. ArXiv e-prints, arXiv:2302.09267, 2024 b

    2024 arXiv

  43. [51]

    Federated learning with adaptive communication compression under dynamic bandwidth and unreliable networks

    Xiongtao Zhang, Xiaomin Zhu, Ji Wang, Hui Yan, Huangke Chen, and Weidong Bao. Federated learning with adaptive communication compression under dynamic bandwidth and unreliable networks. Information Sciences, 540: 0 242--262, 2020

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.