Pith. sign in

REVIEW 3 major objections 4 minor 95 references

Bias-Aware Mislabeling Detection via Decoupled Confident Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read DeCoLe detects mislabeled instances under group-dependent label bias, beating prior confident-learning approaches on hate speech data.

desk verdict Useful problem framing and a rare dual-label evaluation, wrapped around a simple per-group CL; the theory is circular and metrics need fixing, but the empirical claim is credible. read the letter →

arxiv 2507.07216 v2 pith:BXZ4J5KQ submitted 2025-07-09 cs.LG cs.AIcs.DBcs.HC

Yunyi Li , Maria De-Arteaga , Maytal Saar-Tsechansky This is my paper
classification cs.LGcs.AIcs.DBcs.HC
keywords labelbiasmislabelingdetectionconfidentlearninggroup-andclass-conditionalnoisedataqualityhatespeechdecoupledclassifiers
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

The paper proposes DeCoLe (Decoupled Confident Learning), a procedure for finding mislabeled rows in datasets where error rates depend on both the true class and a demographic group, a pattern called label bias. Its central bet is that training one classifier per group and calibrating confident-positive and confident-negative thresholds within each group recovers mislabeled instances that pooled detection methods miss, especially the error type that harms the most affected group. The authors prove that under idealized or bounded-noise predicted probabilities the flagged set matches the true mislabeled set, and they report on synthetic data and a hate speech dataset with dual labels that DeCoLe outperforms Confident Learning, Co-Teaching, and random sampling on recall of mislabels and precision of retained labels. The practical stake is that organizations can audit and relabel only the highest-risk instances instead of paying for gold-standard labels everywhere.

What carries the argument

The load-bearing object is the decoupled threshold pair $(LB_{g_k}, UB_{g_k})$. Instead of pooling errors across groups, DeCoLe fits a model per group and sets each threshold to the group's mean predicted positive probability among observed-positive and observed-negative instances respectively; the confident positive set CPS and confident negative set CNS are then the instances on either side of those thresholds. The identification mechanism is the disagreement between observed label and confident prediction: an observed negative in CPS or an observed positive in CNS is flagged. The theoretical analysis centers on Conditions 1 and 2, which specify when predicted probabilities preserve the rank ordering needed for these thresholds to separate true mislabels from correct labels.

What would settle it

Take any dataset with observed noisy labels, gold-standard labels, and group annotations where label noise is group- and class-conditional, and compute, for each group, the fraction of mislabeled instances whose predicted probability falls on the correct-looking side of $LB_{g_k}$ or $UB_{g_k}$; if that fraction is substantial, DeCoLe's recall collapses even though the noise follows the assumed model. A simpler version: generate noise rates within a group that depend on an additional covariate, and show recall drops sharply.

Watch

Extended reading notes

Core claim

DeCoLe claims that bias-aware mislabeling detection reduces to a set of group-specific confident learning problems. For each group $g_k$, a separate classifier $f_k$ produces out-of-sample predicted probabilities $\hat{p}(\boldsymbol{x}_{g_k})$; DeCoLe computes a lower bound $LB_{g_k}$ as the mean predicted probability of observed positives and an upper bound $UB_{g_k}$ as the mean predicted probability of observed negatives, flags observed negatives above the lower bound and observed positives below the upper bound, and sends those instances to scrutiny. The paper proves the flagged set is a consistent estimator of the truly mislabeled set under an ideal predicted-probability condition and under a per-instance noise condition, and shows empirically that this group-decoupled scheme consistently beats methods that assume class-conditional but group-agnostic noise, with the largest gains for the group bearing the most label bias.

Load-bearing premise

The method assumes that, within each group, the classifier's predicted probabilities rank true positives above false positives and true negatives below false negatives, so the group's mean thresholds slice the mislabeled instances off from the correctly labeled ones; the paper does not independently validate this ranking on data beyond what the models learn.

Editorial extensions

If this is right

  • Datasets with biased labels can be cleaned by flagging only the instances DeCoLe marks, so scarce expert relabeling effort concentrates where errors actually are.
  • DeCoLe gives the largest recall gain for the group suffering the most label bias, while not degrading performance for other groups, which contradicts the common assumption that bias correction forces trade-offs.
  • The method requires only observed labels and group membership; no gold-standard labels are needed at deployment, so it can be applied to datasets where bias is suspected rather than measured.
  • For hate speech moderation, better detection of bias-inducing errors means less harmful content slipping through as non-toxic for groups already disproportionately affected.
  • Because DeCoLe separates the model fitting per group, it also accommodates settings where the same features relate to the label differently across groups (differential subgroup validity).

Reading between the lines

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

  • A testable extension is to replace the mean thresholds with quantile thresholds, which may trade recall for precision when the ranking assumption holds only in the tails; the paper does not explore this.
  • The decoupling principle transfers directly to multi-class labels and intersectional groups, since the threshold logic is per-group and per-class, though the paper only demonstrates binary labels.
  • If the ranking assumption fails, the failure will be most visible in groups with small sample sizes, where the threshold estimates are noisy; auditing those groups first would be a prudent deployment practice.
  • DeCoLe's output could be used as the sampling distribution for active relabeling, prioritizing flagged instances, an integration the paper gestures at but does not evaluate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes DeCoLe (Decoupled Confident Learning), a method for detecting mislabeled instances in datasets where label noise depends on both the latent class and demographic group membership. DeCoLe trains a separate classifier per group, computes group-specific thresholds LB and UB from out-of-sample predicted probabilities, and flags instances whose observed label conflicts with the confident prediction. The authors provide two theoretical conditions under which they claim DeCoLe consistently detects all mislabeled instances, and they evaluate the method on synthetic data and on a hate speech dataset with dual labels (a noisy crowd label and a more expensive gold-standard label), across sexuality, race, and gender groups. The central empirical claim is that DeCoLe consistently outperforms Confident Learning, Co-Teaching, and random sampling in both overall and bias-aware mislabeling detection metrics.

Significance. If the empirical claim holds, the paper would make a useful contribution: it identifies a practically important gap, bias-aware mislabeling detection, and it leverages a rare dataset with noisy labels, gold labels, and demographic annotations. The synthetic experiments are controlled and the real-world evaluation spans multiple demographic attributes, which is more thorough than most work in this area. The paper is also candid about the scarcity of suitable evaluation datasets. However, the theoretical guarantees are much weaker than the framing suggests, and two of the four evaluation metrics appear to be mis-defined. These issues are serious but fixable: the metric definitions can be corrected and the empirical analysis can be supplemented with direct checks of the threshold-separation mechanism.

major comments (3)
  1. [Section 4.1, Eqs. (7) and (8)] The metric definitions in Eqs. (7) and (8) are internally inconsistent. In Eq. (7), the sets D_{\tilde{y}=c,y^*=c|g} and \hat{D}_{\tilde{y}=c,y^*=c|g} denote instances whose observed and gold labels agree, i.e., correctly labeled instances, yet the text describes these as 'mislabeled as c' and uses them to define the recall of bias-inducing errors; the correct sets for false positives/negatives should be D_{\tilde{y}=c,y^*=1-c|g}. In Eq. (8), the denominator is |D_{\tilde{y}=y^*=c|g}|, the number of truly correctly labeled instances of class c, which makes the quantity a recall, not the 'precision of the bias-dominant class in data estimated as correctly labeled'; the denominator should be |\hat{D}_{\tilde{y}=y^*=c|g}|. These errors affect all reported bias-aware metrics, including Tables 4, 5, 8, 9, 12, 13 and Figures 4-7, so the corresponding conclusions about precision and recall gains may not measure what the paper claims.
  2. [Section 3.3, Condition 2 and Theorem 2] Condition 2 defines the per-instance noise distribution with lower and upper endpoints that explicitly contain LB*_{g_k} and UB*_{g_k}. As a result, whenever p(x)>1/2 the perturbed probability is forced to be at least LB*_{g_k}+\epsilon_k, and whenever p(x)<1/2 it is forced to be at most UB*_{g_k}+\epsilon_k. Theorem 2's conclusion therefore follows immediately from the definition of the condition, making the guarantee true by construction rather than a consequence of the learning setup or of the label-noise model. The text's claim that DeCoLe 'remains a consistent estimator' under noisy predicted probabilities is vacuous unless Condition 2 is independently motivated or empirically validated. I recommend reframing this result as a stability condition and adding an explicit empirical check of whether real classifiers approximately satisfy it.
  3. [Section 4.3] The hate speech evaluation never verifies the threshold-separation property on which DeCoLe's mechanism rests: within each group, true positives must have out-of-sample predicted probabilities at or above LB_g and true negatives at or below UB_g. Since the dataset contains the gold labels y*, this can be tested directly, e.g., by reporting, for each group, the fraction of true positives above LB_g and true negatives below UB_g when DeCoLe's thresholds are applied. Without such a check, the reported aggregate recall and precision gains could be driven by threshold calibration or by the metric definitions rather than by the proposed confident-set mechanism, leaving the 'consistently outperforms' claim underdetermined.
minor comments (4)
  1. [Section 4.2.1] The text contains a typo: 'latent good standard label' should be 'latent gold standard label'.
  2. [Algorithm 1, Part 2] The threshold is written as UB^*_{g_k} in one place and UB_{g_k} in another; the notation should be made consistent.
  3. [Section 4.1, Eq. (7)] The phrase 'estimated set of instances that is mislabeled as c, for which the correct label is c' is self-contradictory; if the correct label is c, the instance is not mislabeled.
  4. [General] The paper does not state whether code and data-preprocessing scripts will be released; given the complexity of the real-data pipeline (DistilBERT embeddings, group filtering, cross-validation), a reproducibility statement would be valuable.

Circularity Check

1 steps flagged · score 6.0 of 10

DeCoLe's theoretical consistency guarantee (Theorem 2) is self-definitional: Condition 2 defines the noise distribution using the LB*/UB* thresholds, so the theorem's conclusion holds by construction. The empirical evaluation is independent and not circular.

  1. self definitional [Section 3.3, Condition 2; Appendix B, proof of Theorem 2 (Eqs. 24-34)]
    "The predicted probability ˆ𝑝(𝒙𝑔𝑘) provided by model 𝑓𝑘 is per-instance diffracted if it follows the relationship ˆ 𝑝(𝒙𝑔𝑘) = 𝑝(𝒙𝑔𝑘)+ 𝜖𝒙𝑔𝑘 where the noise term 𝜖𝒙𝑔𝑘 is drawn from the following distribution: 𝜖𝒙𝑔𝑘∼U[ 𝜖𝑘+ LB∗ 𝑔𝑘−𝑝(𝒙𝑔𝑘),𝜖𝑘− LB∗ 𝑔𝑘+𝑝(𝒙𝑔𝑘)] when 𝑝(𝒙𝑔𝑘) > 1/2; and 𝜖𝒙𝑔𝑘∼U[ 𝜖𝑘− UB∗ 𝑔𝑘+𝑝(𝒙𝑔𝑘),𝜖𝑘+ UB∗ 𝑔𝑘−𝑝(𝒙𝑔𝑘)] when 𝑝(𝒙𝑔𝑘) < 1/2."

    The noise distribution is parameterized by LB*_gk and UB*_gk, which are exactly the thresholds DeCoLe uses to define CPS and CNS. For p(x) > 1/2, the support's lower endpoint is eps_k + LB* - p(x), so any draw satisfies phat = p + eps >= eps_k + LB* precisely when p(x) >= LB*; for p(x) < 1/2, the support's upper endpoint is eps_k + UB* - p(x), analogously enforcing the UB* comparison. Theorem 2 then 'proves' that the detected mislabeled set is unchanged from the ideal case by re-deriving these support-imposed inequalities (Eqs. 27-34). The consistency result therefore holds by construction of the noise model, not by any property of classifiers or data; it is a definitional tautology rather than an independently derived guarantee.

full rationale

The paper's empirical contribution is evaluated against external benchmarks (Confident Learning, Co-Teaching, random sampling) using a real hate-speech dataset with gold labels from [47], so the experimental claim that DeCoLe outperforms alternatives is not circular. The label-bias definition is cited to prior work by the same authors, but that citation is not load-bearing in the sense of a uniqueness theorem or an ansatz smuggled in via citation; DeCoLe's mechanism and experiments stand independently of [50,51]. The main circularity is in the theoretical analysis: Condition 2 defines 'per-instance diffracted' noise by placing each epsilon draw on an interval whose endpoints are anchored at LB*_gk and UB*_gk. This makes the key inequalities of Theorem 2 true by definition, so the consistency proof reduces to restating the support constraints. Condition 1, by contrast, is a strong but explicit assumption about ideal predicted probabilities; the theorem derived from it is a conditional mathematical statement rather than a definitional identity. Overall, the theoretical 'principled justification' is partially circular, while the empirical evaluation is self-contained. Hence a score of 6 is appropriate: one central theoretical guarantee reduces by construction, but the main empirical results are not circular.

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

The method has no explicit free parameters beyond the base classifiers, but it relies on two domain assumptions about the noise structure and the informativeness of predicted probabilities. The theoretical conditions in Section 3.3 are ad hoc constructions that guarantee the desired result by definition, which is the main source of circularity burden.

assumptions (4)
  • domain assumption Label noise is group- and class-conditional, with misclassification rates below 0.5 for each group.
    Stated in Section 3.1 and used throughout; the entire method and theory depend on this noise structure.
  • domain assumption Predicted probabilities from a classifier trained on observed labels reflect confidence in the true label (PAC-style identification).
    Invoked in Section 3.2 as the basis for using thresholds on predicted probabilities to build confident sets.
  • ad hoc to paper Condition 1: ideal predicted probabilities equal the group- and class-conditional noise rates and are constant within each class-group cell.
    Defined in Section 3.3; this assumption makes Theorem 1 a near-tautology because the model is assumed to output exactly the noise rate for every instance in a cell.
  • ad hoc to paper Condition 2: noise around ideal probabilities is drawn from a uniform distribution whose bounds are defined in terms of the thresholds LB* and UB*.
    Defined in Section 3.3; the noise interval is constructed so that the threshold inequalities in Theorem 2 hold by construction, making the proof circular.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bias-Aware Mislabeling Detection via Decoupled Confident Learning." pith.science (2026). https://pith.science/paper/BXZ4J5KQ

@misc{pith2026250707216,
  author       = {Pith},
  title        = {Pith review of: Bias-Aware Mislabeling Detection via Decoupled Confident Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BXZ4J5KQ}},
  note         = {Machine review of arXiv:2507.07216}
}
read the original abstract

Reliable data is a cornerstone of modern organizational systems. A notable data integrity challenge stems from label bias, which refers to systematic errors in a label, a covariate that is central to a quantitative analysis, such that its quality differs across social groups. This type of bias has been conceptually and empirically explored and is widely recognized as a pressing issue across critical domains. However, effective methodologies for addressing it remain scarce. In this work, we propose Decoupled Confident Learning (DeCoLe), a principled machine learning based framework specifically designed to detect mislabeled instances in datasets affected by label bias, enabling bias aware mislabelling detection and facilitating data quality improvement. We theoretically justify the effectiveness of DeCoLe and evaluate its performance in the impactful context of hate speech detection, a domain where label bias is a well documented challenge. Empirical results demonstrate that DeCoLe excels at bias aware mislabeling detection, consistently outperforming alternative approaches for label error detection. Our work identifies and addresses the challenge of bias aware mislabeling detection and offers guidance on how DeCoLe can be integrated into organizational data management practices as a powerful tool to enhance data reliability.

Figures

Figures reproduced from arXiv: 2507.07216 by the authors.

Figure 1
Figure 1. Overview of DeCoLe. Given data with label bias, DeCoLe identifies a subset of [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Illustration of DeCoLe’s error identification process. Each panel shows predicted prob [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Dataset generated with group and class-conditional noise and differential subgroup [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of DeCoLe and alternative algorithms (Random, CoT, and CL) for bias-aware [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: Performance of DeCoLe and alternative algorithms across sexuality groups. (a) Overall [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Performance of DeCoLe and competing algorithms in bias-aware mislabeling detection [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: Performance of DeCoLe and competing algorithms in bias-aware mislabeling detection [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

95 extracted references · 74 canonical work pages

  1. [1]

    Pathways for design research on artificial intelligence

    Ahmed Abbasi, Jeffrey Parsons, Gautam Pant, Olivia R Liu Sheng, and Suprateek Sarker. Pathways for design research on artificial intelligence. Inform. Syst. Res., 2024

    2024

  2. [2]

    How management users view information systems

    Carl R Adams. How management users view information systems. Decis. Sci., 6(2):337–345, 1975

    1975

  3. [3]

    Big data, data science, and analytics: The opportunity and challenge for is research

    Ritu Agarwal and Vasant Dhar. Big data, data science, and analytics: The opportunity and challenge for is research. Inform. Syst. Res., 25(3):443–448, 2014

    2014

  4. [4]

    The effect of differential victim crime reporting on predictive policing systems

    Nil-Jana Akpinar, Maria De-Arteaga, and Alexandra Chouldechova. The effect of differential victim crime reporting on predictive policing systems. In Proc. 2021 ACM Conf. Fairness, Accountability, and Transparency, pages 838–849, 2021

    2021

  5. [5]

    Learning from noisy examples

    Dana Angluin and Philip Laird. Learning from noisy examples. Mach. Learn., 2:343–370, 1988

    1988

  6. [6]

    Managing data quality risk in accounting information systems

    Xue Bai, Manuel Nunez, and Jayant R Kalagnanam. Managing data quality risk in accounting information systems. Inf. Syst. Res., 23(2):453–473, 2012

    2012

  7. [7]

    Modeling data and process quality in multi-input, multi-output information systems

    Donald P Ballou and Harold L Pazer. Modeling data and process quality in multi-input, multi-output information systems. Manag. Sci., 31(2):150–162, 1985

    1985

  8. [8]

    Big data’s disparate impact

    Solon Barocas and Andrew D Selbst. Big data’s disparate impact. Calif. L. Rev., 104:671, 2016

    2016

Show all 95 references
  1. [9]

    Active label cleaning for improved dataset quality under resource constraints.Nat

    M ´elanie Bernhardt, Daniel C Castro, Ryutaro Tanno, Anton Schwaighofer, Kerem C Tezcan, Miguel Monteiro, Shruthi Bannur, Matthew P Lungren, Aditya Nori, Ben Glocker, et al. Active label cleaning for improved dataset quality under resource constraints.Nat. Commun., 13(1):1161, 2022

    2022

  2. [10]

    Machine learning in healthcare: Fairness, issues, and challenges

    Margr ´et Vilborg Bjarnad´ottir and David Anderson. Machine learning in healthcare: Fairness, issues, and challenges. In Pushing the Boundaries: Frontiers in Impactful OR/OM Research, pages 64–83. INFORMS, 2020. 29

    2020

  3. [11]

    Language (technology) is power: A critical survey of” bias” in nlp

    Su Lin Blodgett, Solon Barocas, Hal Daum ´e III, and Hanna Wallach. Language (technology) is power: A critical survey of” bias” in nlp. arXiv preprint arXiv:2005.14050, 2020

    2005 arXiv

  4. [12]

    Man is to computer programmer as woman is to homemaker? debiasing word embeddings

    Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Adv. Neural Inf. Process. Syst., 29, 2016

    2016

  5. [13]

    The challenges of data quality and data quality assessment in the big data era

    Li Cai and Yangyong Zhu. The challenges of data quality and data quality assessment in the big data era. Data Sci. J., 14:2–2, 2015

    2015

  6. [14]

    Adaptive sampling strategies to construct equitable training datasets

    William Cai, Ro Encarnacion, Bobbie Chern, Sam Corbett-Davies, Miranda Bogen, Stevie Bergman, and Sharad Goel. Adaptive sampling strategies to construct equitable training datasets. In Proc. 2022 ACM Conf. Fairness, Accountability, and Transparency, pages 1467– 1478, 2022

    2022

  7. [15]

    Semantics derived automatically from language corpora contain human-like biases

    Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183–186, 2017

    2017

  8. [16]

    Internet, social media and online hate speech

    Sergio Andr ´es Casta ˜no-Pulgar´ın, Natalia Su ´arez-Betancur, Luz Magnolia Tilano Vega, and Harvey Mauricio Herrera L ´opez. Internet, social media and online hate speech. systematic review. Aggress. Violent Beh., 58:101608, 2021

    2021

  9. [17]

    Understanding and utilizing deep neural networks trained with noisy labels

    Pengfei Chen, Ben Ben Liao, Guangyong Chen, and Shengyu Zhang. Understanding and utilizing deep neural networks trained with noisy labels. In International conference on machine learning, pages 1062–1070. PMLR, 2019

    2019

  10. [18]

    Modeling spatial and temporal set-based constraints during conceptual database design

    Faiz Currim and Sudha Ram. Modeling spatial and temporal set-based constraints during conceptual database design. Inf. Syst. Res., 23(1):109–128, 2012

    2012

  11. [19]

    Hate speech classifiers learn normative social stereotypes

    Aida Mostafazadeh Davani, Mohammad Atari, Brendan Kennedy, and Morteza Dehghani. Hate speech classifiers learn normative social stereotypes. Trans. Assoc. Comput. Linguist., 11:300–319, 2023

    2023

  12. [20]

    Dealing with disagree- ments: Looking beyond the majority vote in subjective annotations

    Aida Mostafazadeh Davani, Mark D´ıaz, and Vinodkumar Prabhakaran. Dealing with disagree- ments: Looking beyond the majority vote in subjective annotations. Trans. Assoc. Comput. Linguist., 10:92–110, 2022

    2022

  13. [21]

    Competing on analytics

    Thomas H Davenport et al. Competing on analytics. Harv. Bus. Rev., 84(1):98, 2006

    2006

  14. [22]

    Maximum likelihood estimation of observer error-rates using the em algorithm

    Alexander Philip Dawid and Allan M Skene. Maximum likelihood estimation of observer error-rates using the em algorithm. J. R. Stat. Soc. Ser. C Appl. Stat., 28(1):20–28, 1979

    1979

  15. [23]

    Algorithmic fairness in 30 business analytics: Directions for research and practice

    Maria De-Arteaga, Stefan Feuerriegel, and Maytal Saar-Tsechansky. Algorithmic fairness in 30 business analytics: Directions for research and practice. Prod. Oper. Manag., 31(10):3749– 3770, 2022

    2022

  16. [24]

    Reassessing data quality for information products

    Debabrata Dey and Subodha Kumar. Reassessing data quality for information products. Manag. Sci., 56(12):2316–2322, 2010

    2010

  17. [25]

    A checklist to combat cognitive biases in crowdsourcing

    Tim Draws, Alisa Rieger, Oana Inel, Ujwal Gadiraju, and Nava Tintarev. A checklist to combat cognitive biases in crowdsourcing. InProc. AAAI Conf. Hum. Comput. Crowdsource, volume 9, pages 48–59, 2021

    2021

  18. [26]

    Decoupled classifiers for group-fair and efficient machine learning

    Cynthia Dwork, Nicole Immorlica, Adam Tauman Kalai, and Max Leiserson. Decoupled classifiers for group-fair and efficient machine learning. In Proc. 2018 ACM Conf. Fairness, Accountability, and Transparency, pages 119–133. PMLR, 2018

    2018

  19. [27]

    Cognitive biases in crowdsourcing

    Carsten Eickhoff. Cognitive biases in crowdsourcing. In Proc. 11th ACM Int. Conf. Web Search Data Mining, pages 162–170, 2018

    2018

  20. [28]

    The foundations of cost-sensitive learning

    Charles Elkan. The foundations of cost-sensitive learning. In Int. Jt. Conf. Artif. Intell. , volume 17, pages 973–978. Lawrence Erlbaum Associates Ltd, 2001

    2001

  21. [29]

    Fairness evaluation in presence of biased noisy labels

    Riccardo Fogliato, Alexandra Chouldechova, and Max G’Sell. Fairness evaluation in presence of biased noisy labels. In Int. Conf. Artif. Intell. Stat., pages 2325–2336. PMLR, 2020

    2020

  22. [30]

    Classification in the presence of label noise: a survey

    Beno ˆıt Fr´enay and Michel Verleysen. Classification in the presence of label noise: a survey. IEEE Trans. Neural Netw. Learn. Syst., 25(5):845–869, 2013

    2013

  23. [31]

    “un” fair machine learning algorithms

    Runshan Fu, Manmohan Aseri, Param Vir Singh, and Kannan Srinivasan. “un” fair machine learning algorithms. Manage. Sci., 68(6):4173–4195, 2022

    2022

  24. [32]

    Artificial intelligence and algorithmic bias: Source, detection, mitigation, and implications

    Runshan Fu, Yan Huang, and Param Vir Singh. Artificial intelligence and algorithmic bias: Source, detection, mitigation, and implications. In Pushing the Boundaries: Frontiers in Impactful OR/OM Research, pages 39–63. INFORMS, 2020

    2020

  25. [33]

    Do electronic health record systems increase medicare reimbursements? the moderating effect of the recovery audit program

    Kartik K Ganju, Hilal Atasoy, and Paul A Pavlou. Do electronic health record systems increase medicare reimbursements? the moderating effect of the recovery audit program. Manage. Sci., 68(4):2889–2913, 2022

    2022

  26. [34]

    Training deep neural-networks using a noise adap- tation layer

    Jacob Goldberger and Ehud Ben-Reuven. Training deep neural-networks using a noise adap- tation layer. In International conference on learning representations, 2017

    2017

  27. [35]

    Same same, but different: Conditional multi-task learning for demographic-specific toxicity detection

    Soumyajit Gupta, Sooyong Lee, Maria De-Arteaga, and Matthew Lease. Same same, but different: Conditional multi-task learning for demographic-specific toxicity detection. In Proc. ACM Web Conf. 2023, pages 3689–3700, 2023. 31

    2023

  28. [36]

    Co-teaching: Robust training of deep neural networks with extremely noisy labels

    Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Adv. Neural Inf. Process. Syst., 31, 2018

    2018

  29. [37]

    Data quality for data science, predictive analytics, and big data in supply chain management: An introduction to the problem and suggestions for research and applications.Int

    Benjamin T Hazen, Christopher A Boone, Jeremy D Ezell, and L Allison Jones-Farmer. Data quality for data science, predictive analytics, and big data in supply chain management: An introduction to the problem and suggestions for research and applications.Int. J. Prod. Econ., 15...

    2014

  30. [38]

    Racial bias in pain assessment and treatment recommendations, and false beliefs about biological differences between blacks and whites

    Kelly M Hoffman, Sophie Trawalter, Jordan R Axt, and M Norman Oliver. Racial bias in pain assessment and treatment recommendations, and false beliefs about biological differences between blacks and whites. Proc. Natl. Acad. Sci., 113(16):4296–4301, 2016

    2016

  31. [39]

    Crowdsourcing: How the Power of the Crowd is Driving the Future of Business

    Jeff Howe. Crowdsourcing: How the Power of the Crowd is Driving the Future of Business. Random House, 2008

    2008

  32. [40]

    Differential validity of employment tests by race: A comprehensive review and analysis

    John E Hunter, Frank L Schmidt, and Ronda Hunter. Differential validity of employment tests by race: A comprehensive review and analysis. Psychol. Bull., 86(4):721, 1979

    1979

  33. [41]

    Measurement and fairness

    Abigail Z Jacobs and Hanna Wallach. Measurement and fairness. In Proc. 2021 ACM Conf. Fairness, Accountability, and Transparency, pages 375–385, 2021

    2021

  34. [42]

    Emergence of data analytics in the information systems curriculum

    Musa J Jafar, Jeffry Stephen Babb, and Amjad Abdullat. Emergence of data analytics in the information systems curriculum. Inf. Syst. Educ. J., 15(5):22, 2017

    2017

  35. [43]

    A systematic review of hate speech automatic detection using natural language processing

    Md Saroar Jahan and Mourad Oussalah. A systematic review of hate speech automatic detection using natural language processing. Neurocomputing, 546:126232, 2023

    2023

  36. [44]

    Identifying and correcting label bias in machine learning

    Heinrich Jiang and Ofir Nachum. Identifying and correcting label bias in machine learning. In Int. Conf. Artif. Intell. Stat., pages 702–712. PMLR, 2020

    2020

  37. [45]

    Beyond synthetic noise: Deep learning on controlled noisy labels

    Lu Jiang, Di Huang, Mason Liu, and Weilong Yang. Beyond synthetic noise: Deep learning on controlled noisy labels. In Int. Conf. Mach. Learn., pages 4804–4815. PMLR, 2020

    2020

  38. [46]

    Classifying without discriminating

    Faisal Kamiran and Toon Calders. Classifying without discriminating. In Proc. 2nd Int. Conf. Comput. Control Commun., pages 1–6. IEEE, 2009

    2009

  39. [47]

    Constructing interval variables via faceted rasch measurement and multitask deep learning: a hate speech application

    Chris J Kennedy, Geoff Bacon, Alexander Sahn, and Claudia von Vacano. Constructing interval variables via faceted rasch measurement and multitask deep learning: a hate speech application. arXiv preprint arXiv:2009.10277, 2020

    2009 arXiv

  40. [48]

    Demand-aware career path recommendations: A reinforcement learning approach

    Marios Kokkodis and Panagiotis G Ipeirotis. Demand-aware career path recommendations: A reinforcement learning approach. Manag. Sci., 67(7):4362–4383, 2021. 32

    2021

  41. [49]

    On data reliability assessment in accounting information systems

    Ramayya Krishnan, James Peters, Rema Padman, and David Kaplan. On data reliability assessment in accounting information systems. Inf. Syst. Res., 16(3):307–326, 2005

    2005

  42. [50]

    When more data lead us astray: Active data acquisition in the presence of label bias

    Yunyi Li, Maria De-Arteaga, and Maytal Saar-Tsechansky. When more data lead us astray: Active data acquisition in the presence of label bias. In Proc. AAAI Conf. Hum. Comput. Crowdsource, volume 10, pages 133–146, 2022

    2022

  43. [51]

    Label bias: A pervasive and invisibilized problem

    Yunyi Li, Maria De-Arteaga, and Maytal Saar-Tsechansky. Label bias: A pervasive and invisibilized problem. Not. Am. Math. Soc., 71(8):1069–1077, 2024

    2024

  44. [52]

    Detecting and correcting for label shift with black box predictors

    Zachary Lipton, Yu-Xiang Wang, and Alexander Smola. Detecting and correcting for label shift with black box predictors. In International conference on machine learning , pages 3122–3130. PMLR, 2018

    2018

  45. [53]

    Financial statement audits and data breaches

    Lisa Yao Liu. Financial statement audits and data breaches. Manage. Sci., 2024

    2024

  46. [54]

    Data analytics research-informed teaching in a digital technologies curriculum

    Jing Lu. Data analytics research-informed teaching in a digital technologies curriculum. INFORMS Transactions on Education, 20(2):57–72, 2020

    2020

  47. [55]

    Normalized loss functions for deep learning with noisy labels

    Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah Erfani, and James Bailey. Normalized loss functions for deep learning with noisy labels. InInt. Conf. Machine Learning, pages 6543–6553. PMLR, 2020

    2020

  48. [56]

    Using customer analytics to boost corporate performance, 2014

    McKinsey. Using customer analytics to boost corporate performance, 2014

    2014

  49. [57]

    Community standards – hateful conduct

    Meta Platforms, Inc. Community standards – hateful conduct. https://transparency. meta.com/policies/community-standards/hateful-conduct/, 2025. Accessed: April 30, 2025

    2025

  50. [58]

    Targets of online hate speech in context: a comparative digital social science analysis of comments on public facebook pages from romania and hungary

    Radu Mihai Meza, Hanna-Orsolya Vincze, and Andreea Mogos. Targets of online hate speech in context: a comparative digital social science analysis of comments on public facebook pages from romania and hungary. Intersections East Eur. J. Soc. Polit., 4(4), 2018

    2018

  51. [59]

    Prediction-based decisions and fairness: A catalogue of choices, assumptions, and defini- tions

    Shira Mitchell, Eric Potash, Solon Barocas, Alexander D’ Amour, and Kristian Lum. Prediction-based decisions and fairness: A catalogue of choices, assumptions, and defini- tions. arXiv preprint arXiv:1811.07867, 2018

    2018 arXiv

  52. [60]

    On the inequity of predicting a while hoping for b

    Sendhil Mullainathan and Ziad Obermeyer. On the inequity of predicting a while hoping for b. In AEA Papers and Proceedings, volume 111, pages 37–42. American Economic Association 2014 Broadway, Suite 305, Nashville, TN 37203, 2021

    2014

  53. [61]

    Learning with noisy labels

    Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari. Learning with noisy labels. Adv. Neural Inf. Process. Syst., 26, 2013. 33

    2013

  54. [62]

    Ai risk management framework (ai rmf) playbook, 2023

    NIST. Ai risk management framework (ai rmf) playbook, 2023. Accessed: 2025-01-08

    2023

  55. [63]

    Confident learning: Estimating uncertainty in dataset labels

    Curtis Northcutt, Lu Jiang, and Isaac Chuang. Confident learning: Estimating uncertainty in dataset labels. J. Artif. Intell. Res., 70:1373–1411, 2021

    2021

  56. [64]

    Dissecting racial bias in an algorithm used to manage the health of populations

    Ziad Obermeyer, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464):447–453, 2019

    2019

  57. [65]

    Crowdsourcing stereotypes: Linguistic bias in metadata generated via gwap

    Jahna Otterbacher. Crowdsourcing stereotypes: Linguistic bias in metadata generated via gwap. In Proc. 33rd Annu. ACM Conf. Hum. Factors Comput. Syst., pages 1955–1964, 2015

    1955

  58. [66]

    Assessing data quality for information products: impact of selection, projection, and cartesian product

    Amir Parssian, Sumit Sarkar, and Varghese S Jacob. Assessing data quality for information products: impact of selection, projection, and cartesian product. Manag. Sci., 50(7):967–982, 2004

    2004

  59. [67]

    Problem formulation and fairness

    Samir Passi and Solon Barocas. Problem formulation and fairness. In Proc. 2019 ACM Conf. Fairness, Accountability, and Transparency, pages 39–48, 2019

    2019

  60. [68]

    Making deep neural networks robust to label noise: A loss correction approach

    Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In Proc. IEEE Conf. Comput. Vision pattern recognition, pages 1944–1952, 2017

    1944

  61. [69]

    Machine bias, 2016

    ProPublica. Machine bias, 2016. Accessed: 2025-01-28

    2016

  62. [70]

    The risk of racial bias in hate speech detection

    Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A Smith. The risk of racial bias in hate speech detection. In Proc. 57th Annu. Meet. Assoc. Comput. Linguist. , pages 1668–1678, 2019

    2019

  63. [71]

    To explain or to predict? Stat

    Galit Shmueli. To explain or to predict? Stat. Sci., 25(3):289–310, 2010

    2010

  64. [72]

    Meta- weight-net: Learning an explicit mapping for sample weighting

    Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta- weight-net: Learning an explicit mapping for sample weighting. Adv. Neural Inf. Process. Syst., 32, 2019

    2019

  65. [73]

    The interpretation of interaction in contingency tables

    Edward H Simpson. The interpretation of interaction in contingency tables. Journal of the Royal Statistical Society: Series B (Methodological), 13(2):238–241, 1951

    1951

  66. [74]

    Racial bias in pulse oximetry measurement

    Michael W Sjoding, Robert P Dickson, Theodore J Iwashyna, Steven E Gay, and Thomas S Valley. Racial bias in pulse oximetry measurement. N. Engl. J. Med. , 383(25):2477–2478, 2020

    2020

  67. [75]

    Inferring 34 ground truth from subjective labelling of venus images

    Padhraic Smyth, Usama Fayyad, Michael Burl, Pietro Perona, and Pierre Baldi. Inferring 34 ground truth from subjective labelling of venus images. Adv. Neural Inf. Process. Syst. , 7, 1994

    1994

  68. [76]

    Cheap and fast–but is it good? evaluating non-expert annotations for natural language tasks

    Rion Snow, Brendan O’connor, Dan Jurafsky, and Andrew Y Ng. Cheap and fast–but is it good? evaluating non-expert annotations for natural language tasks. In Proc. 2008 Conf. Empir. Methods Nat. Lang. Process., pages 254–263, 2008

    2008

  69. [77]

    Riedl, and Matthew Lease

    Miriah Steiger, Timir J Bharucha, Sukrit Venkatagiri, Martin J. Riedl, and Matthew Lease. The psychological well-being of content moderators: The emotional labor of commercial moderation and avenues for improving support. In Proc. 2021 CHI Conf. Hum. Factors Comput. Syst., CHI...

    2021

  70. [78]

    Training convolutional networks with noisy labels

    Sainbayar Sukhbaatar, Joan Bruna, Manohar Paluri, Lubomir Bourdev, and Rob Fergus. Training convolutional networks with noisy labels. arXiv preprint arXiv:1406.2080, 2014

    2014 arXiv

  71. [79]

    A framework for understanding sources of harm throughout the machine learning life cycle

    Harini Suresh and John Guttag. A framework for understanding sources of harm throughout the machine learning life cycle. In Proc. ACM Conf. Equity Access Algorithms Mechanisms Optimization, pages 1–9, 2021

    2021

  72. [80]

    Improving medical machine learning models with generative balancing for equity and excellence

    Brandon Theodorou, Benjamin Danek, Venkat Tummala, Shivam Pankaj Kumar, Bradley Ma- lin, and Jimeng Sun. Improving medical machine learning models with generative balancing for equity and excellence. npj Digit. Med., 8(1):1–11, 2025

    2025

  73. [81]

    A theory of the learnable

    Leslie G Valiant. A theory of the learnable. Commun. ACM, 27(11):1134–1142, 1984

    1984

  74. [82]

    Learning with symmetric label noise: The importance of being unhinged

    Brendan Van Rooyen, Aditya Menon, and Robert C Williamson. Learning with symmetric label noise: The importance of being unhinged. Advances in neural information processing systems, 28, 2015

    2015

  75. [83]

    Challenges and frontiers in abusive content detection

    Bertie Vidgen, Alex Harris, Dong Nguyen, Rebekah Tromble, Scott Hale, and Helen Margetts. Challenges and frontiers in abusive content detection. In Proc. 3rd Workshop Abuse. Lang. Online. Association for Computational Linguistics, 2019

    2019

  76. [84]

    Anchoring data quality dimensions in ontological founda- tions

    Yair Wand and Richard Y Wang. Anchoring data quality dimensions in ontological founda- tions. Commun. ACM, 39(11):86–95, 1996

    1996

  77. [85]

    Fair classification with group-dependent label noise

    Jialu Wang, Yang Liu, and Caleb Levy. Fair classification with group-dependent label noise. In Proc. 2021 ACM Conf. Fairness, Accountability, and Transparency, pages 526–536, 2021

    2021

  78. [86]

    The multidimensional wisdom of crowds

    Peter Welinder, Steve Branson, Pietro Perona, and Serge Belongie. The multidimensional wisdom of crowds. Adv. Neural Inf. Process. Syst., 23, 2010

    2010

  79. [87]

    Whose 35 vote should count more: Optimal integration of labels from labelers of unknown expertise

    Jacob Whitehill, Ting-fan Wu, Jacob Bergsma, Javier Movellan, and Paul Ruvolo. Whose 35 vote should count more: Optimal integration of labels from labelers of unknown expertise. Adv. Neural Inf. Process. Syst., 22, 2009

    2009

  80. [88]

    Modeling annotator expertise: Learning when everybody knows a bit of something

    Yan Yan, R ´omer Rosales, Glenn Fung, Mark Schmidt, Gerardo Hermosillo, Luca Bogoni, Linda Moy, and Jennifer Dy. Modeling annotator expertise: Learning when everybody knows a bit of something. In Proceedings of the thirteenth Int. Conf. Artif. Intell. Stat. , pages 932–939. JM...

    2010

  81. [89]

    Unlearning bias in language models by partitioning gradients

    Charles Yu, Sullam Jeoung, Anish Kasi, Pengfei Yu, and Heng Ji. Unlearning bias in language models by partitioning gradients. In Find. Assoc. Comput. Linguist., pages 6032–6048, 2023

    2023

  82. [90]

    Risk scores, label bias, and every- thing but the kitchen sink

    Michael Zanger-Tishler, Julian Nyarko, and Sharad Goel. Risk scores, label bias, and every- thing but the kitchen sink. Sci. Adv., 10(13):eadi8411, 2024

    2024

  83. [91]

    Parametrised data sampling for fairness optimisation

    Vladimiro Zelaya, Paolo Missier, and Dennis Prangle. Parametrised data sampling for fairness optimisation. KDD XAI, 2019

    2019

  84. [92]

    Learning fair repre- sentations

    Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi, and Cynthia Dwork. Learning fair repre- sentations. In Int. Conf. Mach. Learn., pages 325–333. PMLR, 2013

    2013

  85. [93]

    Learning from crowdsourced labeled data: a survey

    Jing Zhang, Xindong Wu, and Victor S Sheng. Learning from crowdsourced labeled data: a survey. Artif. Intell. Rev., 46:543–576, 2016

    2016

  86. [94]

    Learning gender-neutral word embeddings

    Jieyu Zhao, Yichao Zhou, Zeyu Li, Wei Wang, and Kai-Wei Chang. Learning gender-neutral word embeddings. arXiv preprint arXiv:1809.01496, 2018. A Additional Data Generation Details for Controlled Empirical Validation We simulate a four-cluster structure based on group and class...

    2018 arXiv

  87. [95]

    LB ∗ 𝑔𝑘 and UB∗ 𝑔𝑘 denote the value of LB𝑔𝑘 and UB𝑔𝑘 under condition 1

    Here, U denotes a uniform distribution, and 𝜖𝑘 = E𝒙𝒈𝒌 [𝜖𝒙𝑔𝑘], where 𝜖𝒙𝑔𝑘 represents the deviation from the ideal predicted probability. LB ∗ 𝑔𝑘 and UB∗ 𝑔𝑘 denote the value of LB𝑔𝑘 and UB𝑔𝑘 under condition 1. Theorem 2 For a biased dataset 𝑫 B(𝒙, ˜𝑦)𝑛, with group- and class-con...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.