Pith. sign in

REVIEW 4 major objections 3 minor 67 references

Effort-Centric Fairness in Lending Decisions

T0 review · 4 major / 3 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Predictive-parity fairness can hide a gender gap in the effort needed to reverse a loan denial, and a tractable regularizer cuts the gap by more than half.

desk verdict Solid framework and exact logistic results, but the headline empirical claim about rejected applicants rests on an explicitly acknowledged proxy; the paper deserves review but needs reframing or real denial data. read the letter →

arxiv 2607.28847 v1 pith:BJKZJAJB submitted 2026-07-30 q-fin.ST q-fin.GNq-fin.RMstat.CO

Shiqi Fang , Zexun Chen , Jake Ansell This is my paper
classification q-fin.STq-fin.GNq-fin.RMstat.CO MSC 91G4062P05
keywords creditscoringalgorithmicfairnessrecourseeffortparitymaskedinequalitycausalstructuralmodellogisticregressionrisk
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that fairness audits based on predictive parity—equal approval rates, equal errors, equal precision—can certify a credit model as fair while it still imposes systematically heavier requirements on rejected female applicants to earn future approval. It defines an applicant's effort as the minimum weighted cost of feasible changes needed to cross the approval boundary, distinguishes independent feature changes from causal shifts that propagate through a structural model, and compares average effort across protected groups. Using matched mortgage data with continuous and discrete features, the authors report that female rejects sit farther from the approval boundary under models that satisfy statistical parity, equalized odds, and positive predictive value parity. They then show that adding effort disparity to the training objective cuts the gap by more than half under feature-independent actions and by over 90% under causal actions, with expected and unexpected loss broadly stable in the first case and higher in the second, while RAROC remains positive.

What carries the argument

The central object is the minimum weighted effort to cross the decision boundary: c*(x) = min over feasible delta of ||delta||_W subject to score >= threshold. For differentiable classifiers the paper derives the local surrogate c_tilde = q / ||W^{-1/2} g||, where q is the score gap and g the input gradient; for causal additive shifts it becomes r_tilde = q / ||W^{-1/2} P^T g|| with P = (I-A)^{-1} the propagation matrix of the learned linear structural causal model, and exact analogues for logistic regression. These closed forms collapse a bi-level recourse problem into a single-level training objective with soft rejection weights, and a displacement lemma bounds resulting changes in portfol

What would settle it

Obtain a lender dataset that includes applicants who were actually denied and tracks their subsequent repayment or reapplication outcomes, compute the average minimum weighted distance to the approval boundary separately for female and male rejects, and check whether the female gap persists; if it disappears or reverses, the masked-inequality claim fails. A second check: re-estimate the causal graph with alternative orientation priors and see whether the causal-effort gap reduction above 90% is stable.

Watch

Extended reading notes

Core claim

The central claim is the existence of masked inequality: a credit model can pass standard predictive-parity checks yet require rejected female applicants to exert greater minimum effort—measured as weighted distance to the approval boundary—than rejected male applicants. The paper further claims that this disparity is actionable: a soft, differentiable effort disparity can be embedded in logistic-regression training, reducing the gap by roughly 60% under feature-independent actions at the largest tested penalty and above 90% under causal additive-shift actions, with finite, quantified changes in expected loss, unexpected loss, and RAROC.

Load-bearing premise

The empirical demonstration uses originated loans only—Section 7.1 states that the matched performance data contain no denied applicants, so low-scoring approved borrowers stand in for real rejects; if actual denied applicants differ systematically from these proxies, the measured gender effort gap and its reduction under regularisation may not carry over to the true rejection population.

Editorial extensions

If this is right

  • If the central claim is right, predictive-parity audits should be supplemented with effort-parity diagnostics to detect unequal routes back into approval.
  • Lenders can train credit models with effort regularisation and reduce the average gender effort gap among rejects by more than half without large predictive losses.
  • Feature-independent effort regularisation keeps expected and unexpected loss broadly stable, while causal regularisation achieves larger fairness gains at the cost of higher losses and lower RAROC.
  • Effort-based training tends to improve equalized-odds and statistical-parity disparities as a by-product, so the two fairness families need not conflict.
  • The same optimization that produces effort parity yields individual-level action plans that tell rejected applicants which concrete feature changes would move them across the approval threshold.

Reading between the lines

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

  • My inference: the effort-gap finding is conditional on the chosen cost weights W and the estimated causal structure A; a different but equally plausible specification could change the size or even the sign of the gap, so audits should report sensitivity to these choices.
  • My inference: masked inequality likely generalizes beyond mortgages—any recurrent yes/no decision with recourse (hiring, insurance, tenancy) could hide effort disparities that outcome-parity checks miss.
  • My inference: the empirical claim about rejected applicants is only proxied by low-scoring originated borrowers, since denied applicants' repayment outcomes are unobserved; a lender's internal data with true rejects would be the direct test of whether the gap transfers.
  • My inference: if effort parity is adopted in regulation, the natural next step is a longitudinal study tracking whether reduced effort gaps actually increase reapplication and approval rates among protected groups.
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

4 major / 3 minor

Summary. The paper develops an effort-centric fairness framework for credit scoring, measuring the minimum weighted cost for a rejected applicant to cross the approval boundary. It distinguishes feature-independent actions from causal shifts under a linear SCM, derives closed-form logistic efforts and local surrogates for differentiable classifiers, embeds an effort-disparity regularizer into training, and proves finite bounds on expected and unexpected loss changes. Using a matched HMDA–Freddie Mac mortgage sample, it reports that female applicants require greater effort than male applicants even under predictive-parity constraints, that feature-independent regularization reduces the gap by roughly 60%, causal regularization by over 90%, and that financial trade-offs are measurable. The mathematical development is careful, but the headline empirical claim rests on an explicitly acknowledged proxy: the matched performance data contain only originated loans, so the 'rejected' group is actually low-scoring originated borrowers.

Significance. If the framework and empirical findings are accepted, the paper makes a useful contribution: it formalizes a fairness dimension that standard parity metrics miss, provides tractable exact/surrogate expressions for effort, and links effort parity to credit-risk and profitability outcomes. The theoretical machinery—weighted Cauchy–Schwarz closed forms, the logistic specialization, the Lipschitz-based risk bounds, and the actionable pathway construction—is a genuine technical strength and appears internally correct under the stated assumptions. The empirical contribution is more fragile: the central claim about rejected female applicants is not directly supported by the data, which contain no denied applicants. The paper's own Section 7.1 caveat is important and should temper the abstract and findings. With careful reframing and additional robustness, the contribution could be publishable; as written, the empirical headline outruns the evidence.

major comments (4)
  1. [§7.1; Abstract and Findings 1–4] The empirical data contain no denied applicants. Section 7.1 states: 'our empirical sample does not include the subsequent repayment outcomes of applicants who were actually denied credit ... the recourse analysis should accordingly be read as measuring effort relative to this approval proxy rather than as observing real denied applicants.' Yet the abstract and Findings 1–4 assert that 'rejected female applicants require greater effort.' The measured effort gap is computed among low-scoring originated borrowers, and selection into origination can differ systematically from the true denial population. This is a load-bearing gap: the headline empirical finding is not established for actual rejected applicants. All abstract and finding statements should be qualified, or the analysis should be supplemented with data or bounds that address selection.
  2. [§4.2 and §7.1, Eq. (18)] The empirical efforts are computed from the exact unconstrained logistic expressions in Eq. (18), but the data include ordinal features (X2, X8) and bounded continuous features, and Assumption 3 restricts the closed forms to continuous actions. For the number of borrowers, Eq. (18) permits fractional changes such as '+0.7 borrowers', which are not feasible. The text acknowledges that if a feasibility constraint binds, Eq. (18) is only an unconstrained benchmark; however, the reported disparities, gap reductions, and regularized models are based on that benchmark, not on the constrained minimal efforts of Definitions 1–2. The empirical claims about effort parity and its mitigation therefore need to be reconciled with feasibility, or explicitly presented as analyses of the unconstrained benchmark.
  3. [§4.2, Eq. (21)] For the logistic model, Eq. (21) gives ΔGC_log(θ) = ΔFI_log(θ)/γ(w). Thus, at any fixed parameter vector, causal effort disparity is a scalar multiple of feature-independent disparity. The large causal reductions in Table 10 at λ = 0.04 (91% vs. 10% for feature-independent) coincide with a large drop in AUC (0.724 → 0.669). The paper should disentangle the effect of the scalar γ(w) from a genuinely different fairness mechanism, and should explain why optimizing the causal objective outperforms optimizing the feature-independent one by such a margin. Without this analysis, the causal parity results are difficult to interpret.
  4. [Table 3] The baseline and SP rows report identical AUC, Accuracy, and F1 (0.724, 0.719, 0.823) with identical standard errors, yet SP is supposed to enforce statistical parity. If the selected SP configuration is numerically indistinguishable from the baseline, the claim that the models 'satisfy' predictive parity is much weaker than it appears. Similarly, EO and PPV report nonzero disparities (0.013 and 0.005) without a stated tolerance. Please provide the parity-constraint configuration actually used, and define a tolerance for what counts as satisfying each criterion.
minor comments (3)
  1. [§7.2, Table 2] Income has mean 113.615 and standard deviation 759.507; despite winsorization at the 99th percentile, the heavy tail may dominate effort calculations. Consider reporting the winsorized distribution and confirming the effort results are robust to alternative transforms.
  2. [§4.3.1, Eq. (23)] The soft disparity in Eq. (23) is an absolute value of a difference of ratios. The paper states it is differentiable almost everywhere and locally Lipschitz, but the absolute value creates a non-differentiability when the two group means cross. This is harmless for Clarke subgradient methods, but the text should note that the objective is non-smooth at that crossing point.
  3. [§7.5, Table 6] The case study reports '∆Number of borrowers X8 +1' for every plan, yet DtB values are nearly identical across plans. It would be helpful to state whether the ordinal rounding is performed via the greedy algorithm of Appendix B and to report which feature drives the marginal effort difference.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: effort values are computed from the fitted boundary rather than used to define it, and no load-bearing self-citation chain is present.

full rationale

The paper's derivation chain is self-consistent and not circular. Effort for logistic regression is defined exactly as c_i = [τ_logit − z_i]_+ / ||W^{−1/2}w|| (Eq. 18), i.e., a deterministic function of the fitted decision boundary and a pre-specified cost matrix W. This effort is then used to construct the fairness regularizer, but the headline empirical finding—that rejected female applicants require greater effort under predictive parity—is measured after training, not used as an input to the coefficients that are then reported. The reduction in effort disparity under regularization is the direct consequence of optimizing the disparity objective, which is a standard in-processing result rather than a circular prediction. Eq. (21) is an internal algebraic relationship between feature-independent and causal effort under a common logistic model and fixed propagation matrix; it does not assume the empirical conclusion. The risk bounds in Theorem 3 are generic Lipschitz/parameter-displacement bounds and do not presuppose the direction or magnitude of the observed risk changes. The paper contains no apparent self-citations, so no self-citation chain is load-bearing. The acknowledged limitation in Section 7.1—that the matched Freddie Mac data contain no actually denied applicants and that the recourse analysis measures effort relative to an approval proxy—is an external-validity and data-availability concern, not a circularity. It weakens the strength of the empirical claim for the true denied population but does not make the derivation depend on its own conclusion. Overall, no step reduces by construction to its inputs.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The central contribution rests on hand-set cost weights, a fixed threshold, an estimated SCM, and the originated-loan proxy; these are inputs or modeling choices, not derived facts. No new physical entities are introduced; 'masked inequality' is a named phenomenon, not a new entity.

free parameters (4)
  • Feature cost weight matrix W = Main config diag(0.2,1e8,1e8,0.2,0.2,1e8,1e8,0.2,0.2); benchmark W in App G diag(1,1e8,1e8,1,1,1e8,1e8,1,1); four other
    Effort and all disparity reductions are defined through the weighted norm ||δ||_W; different W changes who is 'farther' and what a unit of effort means. No sensitivity of the headline gender gap to W is reported for the main result.
  • Decision threshold τ = 0.5
    The approval boundary is fixed at τ=0.5; effort is the distance from the rejected applicant to this boundary, so the measured gap and the regularisation trade-offs depend on τ.
  • Soft-rejection sharpness κ = 10
    Controls how sharply the soft rejection weights approximate the hard indicator in training; affects the fairness objective and its gradients.
  • Structural matrix A of the linear SCM = Estimated from the matched sample via PC-stable + BIC + Bayesian ridge; not listed numerically
    Causal effort and the propagation matrix P=(I-A)^-1 are built from A. If the estimated causal structure is wrong, causal effort parity is miscomputed; this is a fitted input, not an independent ground truth.
assumptions (7)
  • domain assumption Assumption 1: h is differentiable near each rejected applicant and the first-order expansion is accurate for relevant δ.
    Section 4.2; needed for the local surrogates. Exact for logistic regression but only approximate for general differentiable classifiers in finite samples.
  • domain assumption Assumption 2: feasible action sets are star-shaped about zero and the unconstrained linearised minimiser is feasible and local.
    Section 4.2; bridges the closed-form effort to the true constrained effort. This is not verified for individual applicants in the empirical analysis.
  • domain assumption Assumption 3: directly changed features are continuous and W has full rank.
    Section 4.2; the empirical data contain ordinal and nominal features, so the exact logistic efforts in Eq. (18) are only unconstrained benchmarks when discrete constraints bind.
  • domain assumption Linear SCM with acyclicity, causal sufficiency, and additive structural shifts as the intervention semantics.
    Appendix C; defines how P propagates actions. If the linear SCM is misspecified, causal effort parity is not a true causal quantity.
  • domain assumption The matched HMDA-Freddie Mac sample of originated loans stands in for the full applicant pool; no actual denied applicants have observed repayment outcomes.
    Section 7.1; the central empirical claim about 'rejected applicants' is about model-predicted rejects within originated loans, not actual denials.
  • domain assumption Unexpected loss is computed under the Basel ASRF single-factor model with fixed asset correlation ρ.
    Section 5.1; the UL and RAROC results depend on this regulatory capital model.
  • standard math Standard convex analysis and Clarke subdifferential calculus, including weighted Cauchy-Schwarz and Lebourg's mean value theorem.
    Used in Theorems 1-3 and Lemmas 2-3; standard background not proved in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Effort-Centric Fairness in Lending Decisions." pith.science (2026). https://pith.science/paper/BJKZJAJB

@misc{pith2026260728847,
  author       = {Pith},
  title        = {Pith review of: Effort-Centric Fairness in Lending Decisions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BJKZJAJB}},
  note         = {Machine review of arXiv:2607.28847}
}
read the original abstract

Algorithmic credit scoring must satisfy fairness and explanation requirements, yet prevailing predictive-parity criteria assess only outcomes at the decision point. They can therefore overlook whether rejected applicants face unequal burdens in reaching future approval, a phenomenon we call masked inequality. We develop an effort-centric framework that measures an applicant's effort as the minimum weighted cost of feasible changes required to cross the approval boundary. The framework distinguishes feature-independent actions from additive structural shifts that propagate through a causal model and defines parity by comparing average minimum effort across protected groups. We derive tractable local expressions for general differentiable classifiers and exact expressions for logistic regression, embed them in an in-processing fairness objective, and bound changes in portfolio credit risk. The same optimisation yields actionable pathways to approval. Using mortgage data with continuous and discrete features, we find that rejected female applicants require greater effort even when standard predictive-parity criteria are satisfied. Feature-independent regularisation reduces the effort gap by more than 50\% with modest predictive changes. Causal regularisation yields reductions above 90\% at the tested positive penalty weights, but with larger predictive and risk-return trade-offs. Expected and unexpected losses remain broadly stable under feature-independent regularisation and increase under causal regularisation; RAROC declines but remains positive. These results show that effort parity complements predictive fairness by revealing and mitigating hidden barriers to future credit access while making the associated operational trade-offs explicit.

Figures

Figures reproduced from arXiv: 2607.28847 by the authors.

Figure 1
Figure 1. Distribution of effort (distance to boundary) under the SP model. [PITH_FULL_IMAGE:figures/full_fig_p022_1.png] view at source ↗
Figure 2
Figure 2. Trade-off between predictive performance and effort disparity reduction under feature-independent effort [PITH_FULL_IMAGE:figures/full_fig_p023_2.png] view at source ↗
Figure 3
Figure 3. Evolution of effort (distance to boundary) distributions under feature-independent effort parity. [PITH_FULL_IMAGE:figures/full_fig_p024_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Trade-off between credit risk and effort disparity reduction under feature-independent effort parity. [PITH_FULL_IMAGE:figures/full_fig_p025_4.png]
Figure 5
Figure 5. Figure 5: Trade-off between predictive performance, profit, and effort disparity reduction under causal effort [PITH_FULL_IMAGE:figures/full_fig_p026_5.png]
Figure 6
Figure 6. Figure 6: Evolution of effort (distance to boundary) distributions under causal effort parity. [PITH_FULL_IMAGE:figures/full_fig_p026_6.png]
Figure 7
Figure 7. Figure 7: Trade-off between credit risk, profitability, and effort disparity reduction under causal effort parity. [PITH_FULL_IMAGE:figures/full_fig_p027_7.png]
Figure 8
Figure 8. Figure 8: Evolution of predictive parity metrics under feature-independent effort parity. [PITH_FULL_IMAGE:figures/full_fig_p027_8.png]
Figure 9
Figure 9. Figure 9: Evolution of effort distributions under feature-independent effort parity. Each row corresponds to an [PITH_FULL_IMAGE:figures/full_fig_p046_9.png]
Figure 10
Figure 10. Figure 10: Evolution of effort distributions under causal effort parity. Each row corresponds to an independent [PITH_FULL_IMAGE:figures/full_fig_p047_10.png]
Figure 11
Figure 11. Figure 11: Relationship between predictive performance and effort disparity reduction under feature-independent [PITH_FULL_IMAGE:figures/full_fig_p049_11.png]
Figure 12
Figure 12. Figure 12: Relationship between credit risk, profitability, and effort disparity reduction under feature-independent [PITH_FULL_IMAGE:figures/full_fig_p049_12.png]
Figure 13
Figure 13. Figure 13: Relationship between predictive performance and effort disparity reduction under causal effort parity. [PITH_FULL_IMAGE:figures/full_fig_p050_13.png]
Figure 14
Figure 14. Figure 14: Relationship between credit risk, profitability, and effort disparity reduction under causal effort parity. [PITH_FULL_IMAGE:figures/full_fig_p050_14.png]
Figure 15
Figure 15. Figure 15: Relationship between predictive performance and effort disparity reduction under individual effort parity. [PITH_FULL_IMAGE:figures/full_fig_p052_15.png]
Figure 16
Figure 16. Figure 16: Relationship between credit risk, profitability, and effort disparity reduction under individual effort [PITH_FULL_IMAGE:figures/full_fig_p052_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 1 linked inside Pith

  1. [1]

    Journal of the Operational Research Society , volume=

    Transparency, auditability, and explainability of machine learning models in credit scoring , author=. Journal of the Operational Research Society , volume=. 2022 , publisher=

    2022

  2. [2]

    European journal of operational research , volume=

    A comparison of neural networks and linear scoring models in the credit union environment , author=. European journal of operational research , volume=. 1996 , publisher=

    1996

  3. [3]

    Applied Soft Computing , volume=

    Statistical and machine learning models in credit scoring: A systematic literature survey , author=. Applied Soft Computing , volume=. 2020 , publisher=

    2020

  4. [4]

    Computers & operations research , volume=

    Neural network credit scoring models , author=. Computers & operations research , volume=. 2000 , publisher=

    2000

  5. [5]

    Management Science , year=

    Human--algorithmic bias: Source, evolution, and impact , author=. Management Science , year=

  6. [6]

    Management Science , volume=

    Assessing fair lending risks using race/ethnicity proxies , author=. Management Science , volume=. 2018 , publisher=

    2018

  7. [7]

    Management Science , volume=

    Assessing algorithmic fairness with unobserved protected class using data combination , author=. Management Science , volume=. 2022 , publisher=

    2022

  8. [8]

    Building

    Calders, Toon and Kamiran, Faisal and Pechenizkiy, Mykola , month = dec, year =. Building. 2009

    2009

Show all 67 references
  1. [9]

    Operations Research , year=

    Dynamic pricing with fairness constraints , author=. Operations Research , year=

  2. [10]

    Guidance for firms on the fair treatment of vulnerable customers , author=

  3. [11]

    Production and Operations Management , volume=

    Algorithmic fairness in business analytics: Directions for research and practice , author=. Production and Operations Management , volume=. 2022 , publisher=

    2022

  4. [12]

    Management science , volume=

    The role of decision support systems in attenuating racial biases in healthcare delivery , author=. Management science , volume=. 2020 , publisher=

    2020

  5. [13]

    Production and Operations Management , volume=

    Unequal but fair: Incorporating distributive justice in operational allocation models , author=. Production and Operations Management , volume=. 2021 , publisher=

    2021

  6. [14]

    Classifying without discriminating , language =

    Kamiran, Faisal and Calders, Toon , month = feb, year =. Classifying without discriminating , language =. 2009 2nd

    2009

  7. [15]

    Why should i trust you?

    " Why should i trust you?" Explaining the predictions of any classifier , author=. Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining , pages=

  8. [16]

    Management science , volume=

    Algorithmic bias? An empirical study of apparent gender-based discrimination in the display of STEM career ads , author=. Management science , volume=. 2019 , publisher=

    2019

  9. [17]

    European journal of operational research , volume=

    Benchmarking state-of-the-art classification algorithms for credit scoring: An update of research , author=. European journal of operational research , volume=. 2015 , publisher=

    2015

  10. [18]

    ACM Computing Surveys , volume=

    A survey of algorithmic recourse: contrastive explanations and consequential recommendations , author=. ACM Computing Surveys , volume=. 2022 , publisher=

    2022

  11. [19]

    Management Science , volume=

    Price discrimination with fairness constraints , author=. Management Science , volume=. 2022 , publisher=

    2022

  12. [20]

    Advances in neural information processing systems , volume=

    A unified approach to interpreting model predictions , author=. Advances in neural information processing systems , volume=

  13. [21]

    Joint European conference on machine learning and knowledge discovery in databases , pages=

    Fairness-aware classifier with prejudice remover regularizer , author=. Joint European conference on machine learning and knowledge discovery in databases , pages=. 2012 , organization=

    2012

  14. [22]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    On the fairness of causal algorithmic recourse , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  15. [23]

    Proceedings of the Conference on Fairness, Accountability, and Transparency , pages=

    Actionable recourse in linear classification , author=. Proceedings of the Conference on Fairness, Accountability, and Transparency , pages=. 2019 , publisher=

    2019

  16. [24]

    International Conference on Machine Learning , pages=

    On the Long-term Impact of Algorithmic Decision Policies: Effort Unfairness and Feature Segregation through Social Learning , author=. International Conference on Machine Learning , pages=. 2019 , organization=

    2019

  17. [25]

    Proceedings of the Conference on Fairness, Accountability, and Transparency , pages=

    The disparate effects of strategic manipulation , author=. Proceedings of the Conference on Fairness, Accountability, and Transparency , pages=. 2019 , publisher=

    2019

  18. [26]

    arXiv preprint arXiv:1909.03166 , year=

    Equalizing recourse across groups , author=. arXiv preprint arXiv:1909.03166 , year=

    1909 arXiv

  19. [27]

    Advances in Neural Information Processing Systems , volume=

    Learning models for actionable recourse , author=. Advances in Neural Information Processing Systems , volume=

  20. [28]

    Path-specific counterfactual fairness , volume =

    Chiappa, Silvia , month = jul, year =. Path-specific counterfactual fairness , volume =. Proceedings of the

  21. [29]

    Fairness through awareness , isbn =

    Dwork, Cynthia and Hardt, Moritz and Pitassi, Toniann and Reingold, Omer and Zemel, Richard , year =. Fairness through awareness , isbn =. Proceedings of the 3rd

  22. [30]

    European Journal of Operational Research , author =

    Interpretable machine learning for imbalanced credit scoring datasets , volume =. European Journal of Operational Research , author =. 2024 , keywords =

    2024

  23. [31]

    , author=

    Estimating high-dimensional directed acyclic graphs with the PC-algorithm. , author=. Journal of Machine Learning Research , volume=

  24. [32]

    Journal of machine learning research , volume=

    Optimal structure identification with greedy search , author=. Journal of machine learning research , volume=

  25. [33]

    The annals of statistics , pages=

    Estimating the dimension of a model , author=. The annals of statistics , pages=. 1978 , publisher=

    1978

  26. [34]

    Journal of the Royal Statistical Society Series A: Statistics in Society , volume=

    Generalized linear models , author=. Journal of the Royal Statistical Society Series A: Statistics in Society , volume=. 1972 , publisher=

    1972

  27. [35]

    , author=

    Order-independent constraint-based causal structure learning. , author=. J. Mach. Learn. Res. , volume=

  28. [36]

    Proceedings of the 2021 ACM conference on fairness, accountability, and transparency , pages=

    Algorithmic recourse: from counterfactual explanations to interventions , author=. Proceedings of the 2021 ACM conference on fairness, accountability, and transparency , pages=

    2021

  29. [37]

    Management Science , volume=

    “Un” fair machine learning algorithms , author=. Management Science , volume=. 2022 , publisher=

    2022

  30. [38]

    Advances in neural information processing systems , volume=

    Algorithmic recourse under imperfect causal knowledge: a probabilistic approach , author=. Advances in neural information processing systems , volume=

  31. [39]

    Advances in Neural Information Processing Systems , author =

    On fairness and calibration , volume =. Advances in Neural Information Processing Systems , author =

  32. [40]

    European Journal of Operational Research , author =

    Fairness in credit scoring:. European Journal of Operational Research , author =. 2022 , pages =

    2022

  33. [41]

    European Union , year=

    The. European Union , year=

  34. [42]

    Equality of opportunity in supervised learning , language =

    Hardt, Moritz and Price, Eric and Price, Eric and Srebro, Nati , year =. Equality of opportunity in supervised learning , language =. Advances in

  35. [43]

    ACM Computing Surveys , volume=

    Counterfactual explanations and algorithmic recourses for machine learning: A review , author=. ACM Computing Surveys , volume=. 2024 , publisher=

    2024

  36. [44]

    Advances in Neural Information Processing Systems , author =

    Counterfactual fairness , volume =. Advances in Neural Information Processing Systems , author =

  37. [45]

    Elisa and Huang, Lingxiao and Keswani, Vijay and Vishnoi, Nisheeth K

    Celis, L. Elisa and Huang, Lingxiao and Keswani, Vijay and Vishnoi, Nisheeth K. , month = jan, year =. Classification with. Proceedings of the

  38. [46]

    Journal of the operational research society , volume=

    Benchmarking state-of-the-art classification algorithms for credit scoring , author=. Journal of the operational research society , volume=. 2003 , publisher=

    2003

  39. [47]

    Big data , volume=

    Fair prediction with disparate impact: A study of bias in recidivism prediction instruments , author=. Big data , volume=. 2017 , publisher=

    2017

  40. [48]

    The Journal of Finance , volume=

    Predictably unequal? The effects of machine learning on credit markets , author=. The Journal of Finance , volume=. 2022 , publisher=

    2022

  41. [49]

    IEEE Access , volume=

    Model-agnostic counterfactual explanations in credit scoring , author=. IEEE Access , volume=. 2022 , publisher=

    2022

  42. [50]

    European Journal of Operational Research , year=

    Counterfactual explanations for linear optimization , author=. European Journal of Operational Research , year=

  43. [51]

    European Journal of Operational Research , volume=

    Mathematical optimization modelling for group counterfactual explanations , author=. European Journal of Operational Research , volume=. 2024 , publisher=

    2024

  44. [52]

    European Journal of Operational Research , volume=

    Counterfactual analysis and target setting in benchmarking , author=. European Journal of Operational Research , volume=. 2024 , publisher=

    2024

  45. [53]

    European Journal of Operational Research , volume=

    A model-agnostic and data-independent tabu search algorithm to generate counterfactuals for tabular, image, and text data , author=. European Journal of Operational Research , volume=. 2024 , publisher=

    2024

  46. [54]

    Expert Systems with Applications , volume=

    Generating collective counterfactual explanations in score-based classification via mathematical optimization , author=. Expert Systems with Applications , volume=. 2024 , publisher=

    2024

  47. [55]

    Management Science , year=

    The fairness of credit scoring models , author=. Management Science , year=

  48. [56]

    European Journal of Operational Research , volume=

    Explainable AI for operational research: A defining framework, methods, applications, and a research agenda , author=. European Journal of Operational Research , volume=. 2024 , publisher=

    2024

  49. [57]

    Counterfactual explanations without opening the black box: Automated decisions and the GDPR , author=. Harv. JL & Tech. , volume=. 2018 , publisher=

    2018

  50. [58]

    2009 , publisher=

    Causality , author=. 2009 , publisher=

    2009

  51. [59]

    1974 , howpublished =

    Equal Credit Opportunity Act , author =. 1974 , howpublished =

    1974

  52. [60]

    2011 , note =

    12 CFR Part 1002 -- Equal Credit Opportunity Act (Regulation B) , author =. 2011 , note =

    2011

  53. [61]

    2016 , journal =

    Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the Protection of Natural Persons with Regard to the Processing of Personal Data and on the Free Movement of Such Data, and Repealing Directive 95/46/EC (General Data Protection Regulati...

    2016

  54. [62]

    The Review of Financial Studies , volume=

    Role of the community reinvestment act in mortgage supply and the US housing boom , author=. The Review of Financial Studies , volume=. 2020 , publisher=

    2020

  55. [63]

    The Accounting Review , volume=

    Simplifying complex disclosures: Evidence from disclosure regulation in the mortgage markets , author=. The Accounting Review , volume=. 2023 , publisher=

    2023

  56. [64]

    European Journal of Operational Research , volume=

    Evaluating the stability of model explanations in instance-dependent cost-sensitive credit scoring , author=. European Journal of Operational Research , volume=. 2025 , publisher=

    2025

  57. [65]

    European Journal of Operational Research , volume=

    Inherently interpretable machine learning for credit scoring: Optimal classification tree with hyperplane splits , author=. European Journal of Operational Research , volume=. 2025 , publisher=

    2025

  58. [66]

    Journal of Economic Literature , volume=

    The gender wage gap: Extent, trends, and explanations , author=. Journal of Economic Literature , volume=

  59. [67]

    Consumer Financial Protection Circular 2022-03: Adverse action notification requirements in connection with credit decisions based on complex algorithms , author=

    2022

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.