Pith. sign in

REVIEW 2 major objections 8 minor 100 references

Data Heterogeneity Modeling for Trustworthy Machine Learning

T0 review · 2 major / 8 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 claims that ML systems earn trust only by modeling data heterogeneity—latent sub-populations, different generating mechanisms, shift types—at all four stages: collection, training, evaluation, deployment.

desk verdict A well-organized survey that packages existing heterogeneity methods into a pipeline framework; its motivating causal story is the weakest link. read the letter →

arxiv 2506.00969 v1 pith:MTIHYBU4 submitted 2025-06-01 cs.LG cs.AI

Jiashuo Liu , Peng Cui This is my paper
classification cs.LGcs.AI
keywords DataheterogeneityHeterogeneity-awaremachinelearningOut-of-distributiongeneralizationDistributionshiftSub-populationdiscoveryTrustworthyPredictiveModelevaluation
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

Data heterogeneity—the latent sub-populations, varying generating processes, and differing noise levels inside a dataset—is, this paper argues, the overlooked cause of many failures of machine-learning systems in high-stakes settings: poor out-of-distribution generalization, bias against minority groups, and unreliable evaluation. The paper proposes Heterogeneity-Aware Machine Learning (HAML) as a unified paradigm that puts heterogeneity at the center of all four pipeline stages: data collection, model training, model evaluation, and model deployment. It reviews methods that discover sub-populations, quantify predictive heterogeneity, train with explicit or implicit heterogeneity models, evaluate by shift type, and diagnose failures through shift attribution. If the thesis holds, the practical lesson is that characterizing data should come before choosing models and assumptions.

What carries the argument

The load-bearing idea is heterogeneity made quantitative. Predictive heterogeneity (Definition 3.1) is the maximal gain in predictive $\mathcal{V}$-information $I_{\mathcal{V}}(X \to Y)$ achievable by partitioning the data into environments; a large value signals that no single predictive mechanism explains the data. Heterogeneous Risk Minimization (HRM) operationalizes this during training by jointly learning sub-populations and separating stable features from unstable ones, and $\alpha_0$-distributional stability (Definition 4.1) bounds the worst-case KL divergence in $Y|X$ across sub-populations of size at least $\alpha_0$. For evaluation, relative regret measures how much $P(Y|X)$ shifts between training and test, separating $X$-shifts from $Y|X$-shifts; for deployment, feature attribution scores decompose the performance gap into feature subsets. Together these tools carry the argument that heterogeneity can be measured, trained on, and acted upon at every stage.

What would settle it

Take image benchmarks such as DomainBed and tabular benchmarks such as WhyShift, measure their heterogeneity with predictive heterogeneity, and train subgroup-aware models that exploit that measured structure. The central claim would be falsified if, on datasets where heterogeneity is high and correctly modeled, these methods still fail to beat empirical risk minimization or plain DRO—or if the reported failures of DRO and invariant learning persist even when heterogeneity is fully accounted for.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that the gap between theoretically principled robust methods—distributionally robust optimization and invariant learning—and their poor empirical performance is explained by unmodeled data heterogeneity. DRO assumes the test distribution lies in a prescribed ambiguity set around the training distribution, and invariant learning assumes a stable $Y|X$ relationship across environments; real data often violates these assumptions because it contains sub-populations with distinct predictive mechanisms $X \to Y$. The paper presents evidence that explicitly finding and exploiting those sub-populations—via predictive heterogeneity, heterogeneous risk minimization, and $\alpha_0$-distributional stability—improves robustness, fairness, and reliability, and that shift-type-aware evaluation plus post-deployment shift attribution complete the loop. The broad discovery is unificatory: data heterogeneity is the common dual problem behind fairness, bias, and generalization, and it can be addressed systematically across the ML pipeline.

Load-bearing premise

The whole paradigm rests on the claim that the poor real-world results of DRO and invariant learning are caused by unmodeled data heterogeneity rather than by optimization difficulties, benchmark limitations, or other confounds; if that causal reading is wrong, the motivation for the heterogeneity-aware pipeline loses its force.

Editorial extensions

If this is right

  • Data characterization becomes the first step of any ML pipeline: before choosing a robust objective or an invariance assumption, practitioners measure sub-populations and predictive heterogeneity.
  • Out-of-distribution benchmarks must be annotated by shift type, since algorithms designed for $X$-shifts should not be evaluated on data dominated by $Y|X$-shifts.
  • When heterogeneity is present, pooling data and minimizing average risk is wrong; explicit sub-population modeling (HRM and its graph- and recommendation-system adaptations) should improve minority-group and worst-group performance.
  • Model evaluation should go beyond average accuracy to include active stability testing and error slice discovery, so that where a model fails is part of the assessment.
  • After deployment, performance drops should be attributed to specific feature subsets, guiding targeted data collection and group-balancing updates instead of blind fine-tuning.

Reading between the lines

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

  • Beyond the paper: heterogeneity as defined by predictive heterogeneity is relative to the chosen predictive family $\mathcal{V}$, so the same dataset can look homogeneous to one model class and heterogeneous to another; the survey does not draw out this dependence.
  • Beyond the paper: a testable extension is to apply predictive heterogeneity to LLM training corpora and use it to anticipate contamination or domain-gap failures in evaluation, which the paper lists as a future direction but does not develop.
  • Beyond the paper: if the thesis is right, the value chain of AI shifts toward data work—collection, annotation, sub-population labeling, and shift diagnosis—rather than architecture search, with economic consequences the survey only touches on.
  • Beyond the paper: a synthetic benchmark that continuously varies the number and separation of latent sub-populations could test the hypothesis that HAML's advantage over ERM grows with heterogeneity; no such controlled study is reported.
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

2 major / 8 minor

Summary. The paper proposes 'heterogeneity-aware machine learning' (HAML) as a paradigm that integrates data heterogeneity into all four stages of the ML pipeline: data collection, training, evaluation, and deployment. The main body surveys methods and results around predictive heterogeneity, distributional stability, heterogeneity-aware training, shift-aware evaluation, and post-deployment diagnosis, with applications in healthcare, agriculture, finance, graphs, and recommendations. The manuscript argues that traditional model-centric methods such as distributionally robust optimization (DRO) and invariant learning fail in practice largely because they do not analyze data heterogeneity first, and that organizing the field around this principle yields more robust, fair, and reliable systems.

Significance. If the survey's central program is accepted, it would give data scientists and ML practitioners a common vocabulary for connecting data characterization to model design, evaluation, and diagnosis. The paper's strengths are its clear formal statements of predictive heterogeneity (Definition 3.1), distributional stability (Definition 4.1), relative regret (Section 5.1), and feature attribution (Equation 4); its explicit acknowledgment in Section 7 that current demonstrations are small-scale; and its useful organization of many recent works. The main weakness is that the paper's motivating causal claim, that unmodeled data heterogeneity is the primary cause of DRO/IRM failures, is asserted rather than demonstrated, and the case studies are drawn almost entirely from the authors' own prior work without independent validation. As a survey and position paper, the framework is plausible and worth publishing after revision, but the load-bearing motivation needs to be reframed and supported more carefully.

major comments (2)
  1. [Section 2, Figure 2] The causal diagnosis 'This discrepancy primarily arises from a lack of comprehensive analysis of the data characteristics' is load-bearing for the entire HAML paradigm, but the cited evidence does not establish it. In DomainBed [33], GroupDRO and IRM are evaluated with oracle environment labels, i.e., the actual dataset domains, so 'predefined multiple environments are inaccurately specified' cannot explain the failure to beat ERM on PACS, OfficeHome, etc. Separately, the WhyShift analysis [49] in Section 5.1 shows that common tabular benchmarks are dominated by covariate shifts, so DRO underperformance on those benchmarks does not demonstrate that unmodeled Y|X heterogeneity is the cause. I recommend softening this to one plausible cause among several, and adding a concrete refutable prediction such as 'DRO/IRM gains should track a measured heterogeneity quantity,' or including a controlled comparison.
  2. [Section 3.2, Applications 1 and 2 (Figures 3 and 4)] The COVID-19 and crop-yield case studies are presented as evidence that predictive heterogeneity recovers real subgroups, but no uncertainty quantification, statistical tests, or quantitative agreement metrics are provided. Figure 3(a) and (b) report subgroup means without confidence intervals, and Figure 4(b)'s 'strong alignment' with the true crop types is judged visually. Since these results are used to justify the utility of Definition 3.1, the paper should either report numbers such as adjusted Rand index, bootstrap intervals, or significance tests, or explicitly label these as anecdotal illustrations rather than evidence.
minor comments (8)
  1. [Page 1 (ACM template artifact)] The line 'Please use nonacm option or ACM Engage class to enable CC licenses' is a template instruction that should be removed before submission.
  2. [Section 2] The text 'For any close set Z⊂R^d' should read 'closed set'.
  3. [Section 5.1] The text refers to 'Equation (5.1)' for relative regret, but the displayed equation is unnumbered; add an equation number or remove the reference.
  4. [Reference [67]] The author name 'Kun Kunag' is a typo and should be 'Kun Kuang'.
  5. [Acknowledgments] The name 'Jose Blanche' should be 'Jose Blanchet'.
  6. [Definition 4.1] The notation P_{α0}(P) uses P for both the underlying distribution and the family of sub-populations, which is confusing; consider using a different symbol such as Q_{α0}(P) for the set.
  7. [Figure 2 caption] The caption should state whether error bars are shown and which exact settings from [33, Table 4] are used, so readers can verify the claim that GroupDRO and IRM do not improve over ERM.
  8. [Section 7] The conclusion's admission that 'the empirical demonstrations thus far have primarily focused on relatively small-scale or domain-specific applications' should be echoed in the abstract and introduction to avoid overclaiming 'demonstrate the substantial benefits and potential' of HAML.

Circularity Check

0 steps flagged · score 0.0 of 10

Survey-level position paper with no derivation chain; self-citations are descriptive, not load-bearing in a circular sense.

full rationale

This paper is a survey and position piece that organizes prior work under the umbrella of Heterogeneity-Aware Machine Learning (HAML). It does not claim to derive new theoretical results from first principles, so there is no derivation chain in which an input is secretly the output. The core quantities it discusses—predictive heterogeneity (Definition 3.1), alpha0-distributional stability (Definition 4.1), relative regret (Equation 5.1), and feature attribution (Equation 4)—are cited from prior peer-reviewed publications, and the survey explicitly presents them as existing concepts rather than as new derivations. The motivating claim that DRO and invariant learning underperform because of unmodeled data heterogeneity is a causal interpretation of external benchmarks (DomainBed, tabular datasets), not a step that reduces to its own inputs; any weakness in that interpretation is an evidentiary or correctness concern, not circularity. The case studies are summaries of the authors' own prior applications, but they are not fitted parameters renamed as predictions: they are validated against external structure (age distributions, crop types, ICU outcomes) and are falsifiable. No uniqueness theorem is imported from the authors' prior work, no ansatz is smuggled in via citation, and no known result is merely renamed with new coordinates. The paper's self-citations are frequent but descriptive, and the central claim retains independent content from the broader literature it surveys. Hence no significant circularity is identified.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The survey introduces no new fitted parameters or formal entities. It relies on definitions and empirical demonstrations from prior publications, several authored by the present authors, and on a proposed pipeline taxonomy that is assumed rather than derived.

assumptions (4)
  • domain assumption Data can be meaningfully partitioned into environments or subpopulations, and the maximal predictive information gain over such partitions captures real heterogeneity.
    Definition 3.1 defines predictive heterogeneity as a supremum over environment partitions; Section 3 then interprets the discovered subgroups as genuine risk profiles and crop types.
  • domain assumption The cited empirical benchmarks (DomainBed, tabular datasets in WhyShift) faithfully show that DRO and invariant learning underperform ERM in practice.
    Section 2 uses these benchmarks to motivate the entire paradigm, but the survey does not re-run or audit the comparisons.
  • standard math The theoretical results of cited works (finite-sample bounds, strong duality, stability definitions) are correct as stated.
    The survey builds its narrative on results from [7, 49, 52, 53] without reproving or extending them.
  • ad hoc to paper The four-stage pipeline (data collection, training, evaluation, deployment) is a complete and natural decomposition for organizing heterogeneity-aware ML.
    The survey's structure and conclusions depend on this taxonomy, yet no justification is given for why these four stages are the right organizing axes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data Heterogeneity Modeling for Trustworthy Machine Learning." pith.science (2026). https://pith.science/paper/MTIHYBU4

@misc{pith2026250600969,
  author       = {Pith},
  title        = {Pith review of: Data Heterogeneity Modeling for Trustworthy Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MTIHYBU4}},
  note         = {Machine review of arXiv:2506.00969}
}
read the original abstract

Data heterogeneity plays a pivotal role in determining the performance of machine learning (ML) systems. Traditional algorithms, which are typically designed to optimize average performance, often overlook the intrinsic diversity within datasets. This oversight can lead to a myriad of issues, including unreliable decision-making, inadequate generalization across different domains, unfair outcomes, and false scientific inferences. Hence, a nuanced approach to modeling data heterogeneity is essential for the development of dependable, data-driven systems. In this survey paper, we present a thorough exploration of heterogeneity-aware machine learning, a paradigm that systematically integrates considerations of data heterogeneity throughout the entire ML pipeline -- from data collection and model training to model evaluation and deployment. By applying this approach to a variety of critical fields, including healthcare, agriculture, finance, and recommendation systems, we demonstrate the substantial benefits and potential of heterogeneity-aware ML. These applications underscore how a deeper understanding of data diversity can enhance model robustness, fairness, and reliability and help model diagnosis and improvements. Moreover, we delve into future directions and provide research opportunities for the whole data mining community, aiming to promote the development of heterogeneity-aware ML.

Figures

Figures reproduced from arXiv: 2506.00969 by the authors.

Figure 1
Figure 1. Scope of heterogeneity-aware machine learning, which involves the whole machine learning pipeline and connects [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Target accuracy on typical out-of-distribution gen [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Results on the COVID-19 data. (a): The age distri [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Results on the crop yield data. Each region is colored according to its main crop type, and the shade represents the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Case study by Liu et al. [45] that demonstrates better learned sub-populations lead to better generalization perfor￾mances. Figure from [45]. subsequent to data collection. To maximize data utilization and mit￾igate biases during model training, a careful approach to h…
Figure 6
Figure 6. Figure 6: Relative regret in typical benchmarks [20, 29] (left 5 bars) and seven settings designed in WhyShift benchmark [49] (right 7 bars). Figure from [49]. ought to be validated on data with corresponding shift patterns, simply relying on these benchmarks may result in evalu…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

100 extracted references · 58 canonical work pages

  1. [49]

    Jiashuo Liu, Tianyu Wang, Peng Cui, and Hongseok Namkoong. 2024. On the need for a language describing distribution shifts: Illustrations on tabular datasets. Advances in Neural Information Processing Systems36 (2024)

    work page 2024

  2. [33]

    Ishaan Gulrajani and David Lopez-Paz. 2020. In search of lost domain generaliza- tion.arXiv preprint arXiv:2007.01434(2020)

    arXiv 2020

  3. [1]

    Kartik Ahuja, Ethan Caballero, Dinghuai Zhang, Jean-Christophe Gagnon-Audet, Yoshua Bengio, Ioannis Mitliagkas, and Irina Rish. 2021. Invariance principle meets information bottleneck for out-of-distribution generalization.Advances in Neural Information Processing Systems34 (2021), 3438–3450

    2021

  4. [2]

    Kartik Ahuja, Karthikeyan Shanmugam, Kush Varshney, and Amit Dhurandhar

  5. [3]

    Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. 2019. Invariant risk minimization.arXiv preprint arXiv:1907.02893(2019)

    arXiv 2019

  6. [4]

    2007.Stochastic simulation: algorithms and analysis

    Søren Asmussen and Peter W Glynn. 2007.Stochastic simulation: algorithms and analysis. Vol. 57. Springer

    2007

  7. [5]

    Susan Athey, Julie Tibshirani, and Stefan Wager. 2019. Generalized random forests. (2019)

    2019

  8. [6]

    Pedro Baqui, Ioana Bica, Valerio Marra, Ari Ercole, and Mihaela Van Der Schaar

Show all 100 references
  1. [7]

    Jose Blanchet, Peng Cui, Jiajin Li, and Jiashuo Liu. 2024. Stability evaluation through distributional perturbation analysis. InForty-first International Confer- ence on Machine Learning

    2024

  2. [8]

    Ethnic and regional variation in hospital mortality from COVID-19 in Brazil.medRxiv(2020), 2020–05

    2020

  3. [9]

    Jose Blanchet, Yang Kang, Karthyek Murthy, and Fan Zhang. 2019. Data-driven optimal transport cost selection for distributionally robust optimization. In2019 winter simulation conference (WSC). IEEE, 3740–3751

    2019

  4. [10]

    Jose Blanchet, Yang Kang, and Karthyek Murthy. 2019. Robust Wasserstein profile inference and applications to machine learning.Journal of Applied Probability 56, 3 (2019), 830–857

    2019

  5. [11]

    Tiffany Tianhui Cai, Hongseok Namkoong, and Steve Yadlowsky. 2023. Diagnos- ing model performance under distribution shift.arXiv preprint arXiv:2303.02011 (2023)

    2023 arXiv

  6. [12]

    Jose Blanchet, Jiajin Li, Sirui Lin, and Xuhui Zhang. 2024. Distributionally robust optimization and robust statistics.arXiv preprint arXiv:2401.14655(2024)

    2024 arXiv

  7. [13]

    Wentao Chen, Lizhe Zhang, Li Zhong, Letian Peng, Zilong Wang, and Jingbo Shang. 2025. Memorize or Generalize? Evaluating LLM Code Generation with Evolved Questions.arXiv e-prints(2025), arXiv–2503

    2025

  8. [14]

    Dan A Calian, Gregory Farquhar, Iurii Kemaev, Luisa M Zintgraf, Matteo Hessel, Jeremy Shar, Junhyuk Oh, András György, Tom Schaul, Jeffrey Dean, et al. 2025. DataRater: Meta-Learned Dataset Curation.arXiv preprint arXiv:2505.17895 (2025)

    2025

  9. [15]

    Yongqiang Chen, Kaiwen Zhou, Yatao Bian, Binghui Xie, Bingzhe Wu, Yonggang Zhang, Kaili Ma, Han Yang, Peilin Zhao, Bo Han, et al . 2022. Pareto invari- ant risk minimization: Towards mitigating the optimization dilemma in out-of- distribution generalization.arXiv preprint arXi...

    2022 arXiv

  10. [16]

    Yongqiang Chen, Yatao Bian, Kaiwen Zhou, Binghui Xie, Bo Han, and James Cheng. 2024. Does invariant graph learning via environment augmentation learn invariance?Advances in Neural Information Processing Systems36 (2024)

    2024

  11. [17]

    Issa J Dahabreh, Rodney Hayward, and David M Kent. 2016. Using group data to treat individuals: understanding heterogeneous treatment effects in the age of precision medicine and patient-centred evidence.International journal of epidemiology45, 6 (2016), 2184–2193

    2016

  12. [18]

    Elliot Creager, Jörn-Henrik Jacobsen, and Richard Zemel. 2021. Environment inference for invariant learning. InInternational Conference on Machine Learning. PMLR, 2189–2200

    2021

  13. [19]

    Xiaoyu Du, Zike Wu, Fuli Feng, Xiangnan He, and Jinhui Tang. 2022. In- variant Representation Learning for Multimedia Recommendation. InPro- ceedings of the 30th ACM International Conference on Multimedia(<conf-loc>, <city>Lisboa</city>, <country>Portugal</country>, </conf-loc...

    2022

  14. [20]

    Frances Ding, Moritz Hardt, John Miller, and Ludwig Schmidt. 2021. Retiring adult: New datasets for fair machine learning.Advances in neural information processing systems34 (2021), 6478–6490

    2021

  15. [21]

    John Duchi, Tatsunori Hashimoto, and Hongseok Namkoong. 2023. Distribution- ally robust losses for latent covariate mixtures.Operations Research71, 2 (2023), 649–664

    2023

  16. [22]

    Dheeru Dua and Casey Graff. 2017. UCI Machine Learning Repository. http: //archive.ics.uci.edu/ml

    2017

  17. [23]

    Kathleen C Engel and Patricia A McCoy. 2008. From credit denial to predatory lending: The challenge of sustaining minority homeownership. InSegregation. Routledge, 97–140

    2008

  18. [24]

    John C Duchi and Hongseok Namkoong. 2021. Learning models with uniform performance via distributionally robust optimization.The Annals of Statistics49, 3 (2021), 1378–1406

    2021

  19. [25]

    Jianqing Fan, Fang Han, and Han Liu. 2014. Challenges of big data analysis. National science review1, 2 (2014), 293–314

    2014

  20. [26]

    Sabri Eyuboglu, Maya Varma, Khaled Kamal Saab, Jean-Benoit Delbrouck, Christo- pher Lee-Messer, Jared Dunnmon, James Zou, and Christopher Re. 2022. Domino: Discovering Systematic Errors with Cross-Modal Embeddings. InInternational Conference on Learning Representations

    2022

  21. [27]

    Shaohua Fan, Renzhe Xu, Qian Dong, Yue He, Cheng Chang, and Peng Cui. 2024. Stable Cox regression for survival analysis under distribution shifts.Nature Machine Intelligence6, 12 (2024), 1525–1541

    2024

  22. [28]

    Jianqing Fan, Ricardo Masini, and Marcelo C Medeiros. 2022. Do we exploit all in- formation for counterfactual analysis? Benefits of factor models and idiosyncratic correction.J. Amer. Statist. Assoc.117, 538 (2022), 574–590

    2022

  23. [29]

    Josh Gardner, Zoran Popović, and Ludwig Schmidt. 2022. Subgroup Robustness Grows On Trees: An Empirical Baseline Investigation. arXiv:2211.12703 [cs]

    2022 arXiv

  24. [30]

    Yaroslav Ganin and Victor Lempitsky. 2015. Unsupervised domain adaptation by backpropagation. InInternational conference on machine learning. PMLR, 1180– 1189

    2015

  25. [31]

    Roger Grosse, Juhan Bae, Cem Anil, Nelson Elhage, Alex Tamkin, Amirhossein Tajdini, Benoit Steiner, Dustin Li, Esin Durmus, Ethan Perez, et al. 2023. Studying large language model generalization with influence functions.arXiv preprint arXiv:2308.03296(2023)

    2023 arXiv

  26. [32]

    Shantanu Ghosh, Rayan Syed, Chenyu Wang, Clare B Poynton, Shyam Visweswaran, and Kayhan Batmanghelich. 2024. LADDER: Language Driven Slice Discovery and Error Rectification.arXiv preprint arXiv:2408.07832(2024)

    2024 arXiv

  27. [34]

    Shurui Gui, Meng Liu, Xiner Li, Youzhi Luo, and Shuiwang Ji. 2024. Joint learning of label and environment causal independence for graph out-of-distribution generalization.Advances in Neural Information Processing Systems36 (2024)

    2024

  28. [35]

    Bo-Wei Huang, Keng-Te Liao, Chang-Sheng Kao, and Shou-De Lin. 2022. Envi- ronment Diversification with Multi-head Neural Network for Invariant Learning. Advances in Neural Information Processing Systems35 (2022), 915–927

    2022

  29. [36]

    Moritz Hardt, Eric Price, and Nati Srebro. 2016. Equality of opportunity in supervised learning.Advances in neural information processing systems29 (2016)

    2016

  30. [37]

    Nari Johnson, Ángel Alexander Cabrera, Gregory Plumb, and Ameet Talwalkar

  31. [38]

    Sérgio Jesus, José Pombal, Duarte Alves, André Cruz, Pedro Saleiro, Rita Ribeiro, João Gama, and Pedro Bizarro. 2022. Turning the tables: Biased, imbalanced, dynamic tabular datasets for ml evaluation.Advances in Neural Information Processing Systems35 (2022), 33563–33575

    2022

  32. [39]

    Pang Wei Koh and Percy Liang. 2017. Understanding black-box predictions via influence functions. InInternational conference on machine learning. PMLR, 1885–1894

    2017

  33. [40]

    Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. 2021. Wilds: A benchmark of in-the-wild distribution shifts. InInternational Conference on Machine Learni...

    2021

  34. [41]

    David M Kent, Ewout Steyerberg, and David Van Klaveren. 2018. Personalized evidence based medicine: predictive approaches to heterogeneous treatment effects.Bmj363 (2018)

    2018

  35. [42]

    Tian Li, Ahmad Beirami, Maziar Sanjabi, and Virginia Smith. 2023. On tilted losses in machine learning: Theory and applications.Journal of Machine Learning Research24, 142 (2023), 1–79

    2023

  36. [43]

    Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. 2021. Just train twice: Improving group robustness without training group information. InInternational Conference on Machine Learning. PMLR, 6781–6792

    2021

  37. [44]

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. 2017. Deeper, broader and artier domain generalization. InProceedings of the IEEE international conference on computer vision. 5542–5550

    2017

  38. [45]

    Jiashuo Liu, Zheyuan Hu, Peng Cui, Bo Li, and Zheyan Shen. 2021. Kernelized heterogeneous risk minimization.NeurIPS(2021)

    2021

  39. [46]

    Jiashuo Liu, Nabeel Seedat, Peng Cui, and Mihaela van der Schaar. 2025. Go- ing Beyond Static: Understanding Shifts with Time-Series Attribution. InThe Thirteenth International Conference on Learning Representations

    2025

  40. [47]

    Jiashuo Liu, Zheyuan Hu, Peng Cui, Bo Li, and Zheyan Shen. 2021. Heterogeneous risk minimization. InInternational Conference on Machine Learning. PMLR, 6804– 6814. KDD ’25, August 3–7, 2025, Toronto, ON, Canada Jiashuo Liu and Peng Cui

    2021

  41. [48]

    Jiashuo Liu, Zheyan Shen, Peng Cui, Linjun Zhou, Kun Kuang, Bo Li, and Yishi Lin. 2021. Stable adversarial learning under distributional shifts. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 8662–8670

    2021

  42. [50]

    Jiashuo Liu, Zheyan Shen, Peng Cui, Linjun Zhou, Kun Kuang, and Bo Li. 2022. Distributionally robust learning with stable adversarial training.IEEE Transac- tions on Knowledge and Data Engineering(2022)

    2022

  43. [51]

    Jiashuo Liu, Jiayun Wu, Bo Li, and Peng Cui. 2022. Distributionally Robust Optimization with Data Geometry. InAdvances in Neural Information Processing Systems

    2022

  44. [52]

    Jiashuo Liu, Jiayun Wu, Jie Peng, Xiaoyu Wu, Yang Zheng, Bo Li, and Peng Cui. 2024. Enhancing Distributional Stability among Sub-populations.AISTATS (2024)

    2024

  45. [53]

    Jiashuo Liu, Tianyu Wang, Henry Lam, Hongseok Namkoong, and Jose Blanchet

  46. [54]

    Jiashuo Liu, Jiayun Wu, Tianyu Wang, Hao Zou, and Peng Cui. 2023. Geometry- Calibrated DRO: Combating Over-Pessimism with Free Energy Implications. InNeurIPS 2023 Workshop on Distribution Shifts: New Frontiers with Foundation Models

    2023

  47. [55]

    David B Lobell, Marshall B Burke, Claudia Tebaldi, Michael D Mastrandrea, Walter P Falcon, and Rosamond L Naylor. 2008. Prioritizing climate change adaptation needs for food security in 2030.Science319, 5863 (2008), 607–610

    2008

  48. [56]

    John P Miller, Rohan Taori, Aditi Raghunathan, Shiori Sagawa, Pang Wei Koh, Vaishaal Shankar, Percy Liang, Yair Carmon, and Ludwig Schmidt. 2021. Ac- curacy on the line: on the strong correlation between out-of-distribution and in-distribution generalization. InInternational c...

    2021

  49. [57]

    Jiashuo Liu, Jiayun Wu, Renjie Pi, Renzhe Xu, Xingxuan Zhang, Bo Li, and Peng Cui. 2023. Measure the predictive heterogeneity.ICLR(2023)

    2023

  50. [58]

    Ziad Obermeyer, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. 2019. Dissecting racial bias in an algorithm used to manage the health of populations. Science366, 6464 (2019), 447–453

    2019

  51. [59]

    Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. 2020. Identifying mislabeled data using the area under the margin ranking.Advances in Neural Information Processing Systems33 (2020), 17044–17056

    2020

  52. [60]

    Michael Roberts, Derek Driggs, Matthew Thorpe, Julian Gilbey, Michael Yeung, Stephan Ursprung, Angelica I Aviles-Rivero, Christian Etmann, Cathal McCague, Lucian Beer, et al. 2021. Common pitfalls and recommendations for using machine learning to detect and prognosticate for C...

    2021

  53. [61]

    Peyman Mohajerin Esfahani and Daniel Kuhn. 2018. Data-driven distributionally robust optimization using the Wasserstein metric: Performance guarantees and tractable reformulations.Mathematical Programming171, 1 (2018), 115–166

    2018

  54. [62]

    Everyone wants to do the model work, not the data work

    Nithya Sambasivan, Shivani Kapania, Hannah Highfill, Diana Akrong, Praveen Paritosh, and Lora M Aroyo. 2021. “Everyone wants to do the model work, not the data work”: Data Cascades in High-Stakes AI. Inproceedings of the 2021 CHI Conference on Human Factors in Computing Systems. 1–15

    2021

  55. [63]

    Katia Savchuk. [n. d.]. Big Data and Racial Bias: Can That Ghost Be Removed from the Machine? https://www.gsb.stanford.edu/insights/big-data-racial-bias- can-ghost-be-removed-machine. Accessed: 2019-10-28

    2019

  56. [64]

    Nabeel Seedat, Jonathan Crabbé, Ioana Bica, and Mihaela van der Schaar. 2022. Data-iq: Characterizing subgroups with heterogeneous outcomes in tabular data. Advances in Neural Information Processing Systems35 (2022), 23660–23674

    2022

  57. [65]

    Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. 2019. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization.arXiv preprint arXiv:1911.08731 (2019)

    2019 arXiv

  58. [66]

    Nabeel Seedat, Nicolas Huynh, Fergus Imrie, and Mihaela van der Schaar. 2024. You can’t handle the (dirty) truth: Data-centric insights improve pseudo-labeling. Journal of Data-centric Machine Learning Research(2024)

    2024

  59. [67]

    Zheyan Shen, Peng Cui, Tong Zhang, and Kun Kunag. 2020. Stable learning via sample reweighting. InProceedings of the AAAI Conference on Artificial Intelli- gence, Vol. 34. 5692–5699

    2020

  60. [68]

    Zheyan Shen, Jiashuo Liu, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. 2021. Towards out-of-distribution generalization: A survey.arXiv preprint arXiv:2108.13624(2021)

    2021 arXiv

  61. [69]

    Nabeel Seedat, Jonathan Crabbé, Zhaozhi Qian, and Mihaela van der Schaar

  62. [70]

    Advances in Neural Information Processing Systems36 (2023), 74995–75008

    Triage: Characterizing and auditing training data for improved regression. Advances in Neural Information Processing Systems36 (2023), 74995–75008

    2023

  63. [71]

    Matthew Staib and Stefanie Jegelka. 2019. Distributionally robust optimization and generalization in kernel methods.Advances in Neural Information Processing Systems32 (2019)

    2019

  64. [72]

    Smith, and Yejin Choi

    Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Han- naneh Hajishirzi, Noah A. Smith, and Yejin Choi. 2020. Dataset Cartogra- phy: Mapping and Diagnosing Datasets with Training Dynamics. InProceed- ings of the 2020 Conference on Empirical Methods in Natural La...

    2020

  65. [73]

    Nikolaj Thams, Michael Oberst, and David Sontag. 2022. Evaluating robustness to dataset shift via parametric robustness sets.Advances in Neural Information Processing Systems35 (2022), 16877–16889

    2022

  66. [74]

    Hidetoshi Shimodaira. 2000. Improving predictive inference under covariate shift by weighting the log-likelihood function.Journal of statistical planning and inference90, 2 (2000), 227–244

    2000

  67. [75]

    Harvineet Singh, Fan Xia, Adarsh Subbaswamy, Alexej Gossmann, and Jean Feng

  68. [76]

    Shikha Verma. 2019. Weapons of math destruction: how big data increases inequality and threatens democracy.Vikalpa44, 2 (2019), 97–98

    2019

  69. [77]

    Stefan Wager and Susan Athey. 2018. Estimation and inference of heterogeneous treatment effects using random forests.J. Amer. Statist. Assoc.113, 523 (2018), 1228–1242

    2018

  70. [78]

    Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. 2022. Generalizing to unseen domains: A survey on domain generalization.IEEE Transactions on Knowledge and Data Engineering(2022)

    2022

  71. [79]

    Zimu Wang, Yue He, Jiashuo Liu, Wenchao Zou, Philip S Yu, and Peng Cui

  72. [80]

    Tipton, J

    E. Tipton, J. Spybrook, K. G. Fitzgerald, Q. Wang, and C. Davidson. 2020. Toward a System of Evidence for All: Current Practices and Future Opportunities in 37 Randomized Trials.Educational Researcher(2020), 0013189X2096068

    2020

  73. [81]

    Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J Gordon. 2019. An Empirical Study of Example Forgetting during Deep Neural Network Learning. InInternational Conference on Learning Representations

    2019

  74. [82]

    Laure Wynants, Ben Van Calster, Gary S Collins, Richard D Riley, Georg Heinze, Ewoud Schuit, Marc MJ Bonten, Darren L Dahly, Johanna A Damen, Thomas PA Debray, et al. 2020. Prediction models for diagnosis and prognosis of covid-19: systematic review and critical appraisal.bmj3...

    2020

  75. [83]

    Chunqiu Steven Xia, Yinlin Deng, and LINGMING ZHANG. 2024. Top Leader- board Ranking= Top Coding Proficiency, Always? EvoEval: Evolving Coding Benchmarks via LLM. InFirst Conference on Language Modeling

    2024

  76. [84]

    Kelin Xu, Liping Zhu, and Jianqing Fan. 2022. Distributed sufficient dimension reduction for heterogeneous massive data.Statistica Sinica32 (2022), 2455–2476

    2022

  77. [85]

    Renzhe Xu, Kang Wang, and Bo Li. 2025. Heterogeneous Data Game: Charac- terizing the Model Competition Across Multiple Data Sources.International Conference on Machine Learning

    2025

  78. [86]

    Yuzhe Yang, Haoran Zhang, Dina Katabi, and Marzyeh Ghassemi. 2023. Change is hard: a closer look at subpopulation shift. InProceedings of the 40th International Conference on Machine Learning. 39584–39622

    2023

  79. [87]

    Colin White, Samuel Dooley, Manley Roberts, Arka Pal, Benjamin Feuer, Sid- dhartha Jain, Ravid Shwartz-Ziv, Neel Jain, Khalid Saifullah, Sreemanti Dey, et al

  80. [88]

    InThe Thirteenth International Conference on Learning Representations

    LiveBench: A Challenging, Contamination-Free LLM Benchmark. InThe Thirteenth International Conference on Learning Representations

  81. [89]

    Qitian Wu, Hengrui Zhang, Junchi Yan, and David Wipf. 2021. Handling Distri- bution Shifts on Graphs: An Invariance Perspective. InInternational Conference on Learning Representations

    2021

  82. [90]

    Weihuang Zheng, Jiashuo Liu, Jiaxing Li, Jiayun Wu, Peng Cui, and Youyong Kong. 2024. Topology-Aware Dynamic Reweighting for Distribution Shifts on Graph. InInternational Conference on Machine Learning

    2024

  83. [91]

    Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. 2022. Domain generalization: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 4 (2022), 4396–4415

    2022

  84. [92]

    Yang Zhou, Lirong Xue, Zhengyu Shi, Libo Wu, and Jianqing Fan. 2022. Measuring housing vitality from multi-source big data and machine learning.J. Amer. Statist. Assoc.117, 539 (2022), 1045–1059

    2022

  85. [95]

    Jinsung Yoon, Sercan Arik, and Tomas Pfister. 2020. Data valuation using re- inforcement learning. InInternational Conference on Machine Learning. PMLR, 10842–10851

    2020

  86. [96]

    Han Yu, Jiashuo Liu, Hao Zou, Renzhe Xu, Yue He, Xingxuan Zhang, and Peng Cui. 2025. Error Slice Discovery via Manifold Compactness.arXiv preprint arXiv:2501.19032(2025)

    2025

  87. [97]

    Xingxuan Zhang, Linjun Zhou, Renzhe Xu, Peng Cui, Zheyan Shen, and Haoxin Liu. 2023. Nico++: Towards better benchmarking for domain generalization. CVPR(2023)

    2023

  88. [2020]

    InInternational Conference on Machine Learning

    Invariant risk minimization games. InInternational Conference on Machine Learning. PMLR, 145–155

  89. [2022]

    InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Invariant preference learning for general debiasing in recommendation. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1969–1978

    1969

  90. [2023]

    InProceedings of the AAAI Conference on Human Computation and Crowdsourcing, Vol

    Where does my model underperform? a human evaluation of slice discovery algorithms. InProceedings of the AAAI Conference on Human Computation and Crowdsourcing, Vol. 11. 65–76

  91. [2024]

    Advances in Neural Information Processing Systems37 (2024), 128516–128555

    A hierarchical decomposition for explaining ML performance discrepancies. Advances in Neural Information Processing Systems37 (2024), 128516–128555

    2024

  92. [2025]

    arXiv:2505.23565 [cs.LG] https://arxiv.org/abs/2505.23565

    DRO: A Python Library for Distributionally Robust Optimization in Ma- chine Learning. arXiv:2505.23565 [cs.LG] https://arxiv.org/abs/2505.23565

    arXiv

Pith tools

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