Pith. sign in

REVIEW 3 major objections 6 minor 45 references

Accurate Ensembles, Fragile Narratives: Multi-Scale Stacking and a Fidelity Audit of LLM-Generated Explanations for Credit Risk

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A constrained LLM credit explanation inverted the risk directions it was handed, and the paper shows why prompt-side controls alone fail.

desk verdict A genuinely honest negative result about LLM-generated credit explanations, but the headline sign-inversion count is inflated and the audit's ground truth is shaky. read the letter →

arxiv 2608.08126 v1 pith:327WZ6NP submitted 2026-08-08 cs.LG cs.CL

Gregorius Reynaldi Pratama , Kuo-Kun Tseng This is my paper · ORCID
classification cs.LGcs.CL
keywords creditriskstackingensembleLLM-generatedexplanationsexplanationfidelitySHAPLIMEattributioninstabilitymodelcalibration
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 builds a credit-risk explanation pipeline of the kind now common in the literature—a high-accuracy stacking ensemble, SHAP and LIME attributions, and a constrained language model that writes the applicant-facing rationale—and then asks whether the generated narrative matches the attributions it was given. It finds that the predictive gains are real but small at the deployed threshold, while the narrative layer fails: in a documented case the model said age, income, and venture intent increase risk when the supplied attributions scored all three as risk-reducing, omitted the dominant driver, and introduced a feature never provided. The paper's central conclusion is that constrained prompting is necessary but not sufficient, and that grounding must be verified after generation with cheap membership, direction, and coverage checks. A sympathetic reader should take this as an existence proof that the standard input-side control architecture can produce fluent, sign-inverted rationales.

What carries the argument

The load-bearing object is the dual-layer explanation pipeline and the audit built to inspect it. Layer one fuses min-max-normalized SHAP and LIME scores into a single score per feature, truncates to the top four drivers, and hands them with signed impacts to a 2.7-billion-parameter decoder-only language model under a deliberately constrained prompt with greedy decoding. Layer two is the fidelity audit introduced by the paper: membership (every feature named in the narrative must be in the supplied driver set), direction (the narrative's asserted risk direction must match the sign of the fused attribution), and coverage (the top driver must be named). The audit is what carries the argument, because it turns "the narrative reads well" into checkable propositions that can contradict the evidence.

What would settle it

Re-score the audited case with SHAP only and with LIME only as the reference, regenerate the narrative under the same prompt, and check whether the narrative matches either reference; if it matches one, the reported sign inversion is an artifact of the fusion step rather than a failure of the language model to follow supplied signs. Alternatively, run the paper's three audit checks on a stratified sample of several hundred test instances; violation rates near zero across that sample would undercut the claim that this failure is a systematic property of the architecture.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is asymmetric. The multi-scale stacking ensemble reaches test ROC-AUC 0.9539 and PR-AUC 0.9137, beating the best single model by $\Delta\text{AUC} = 0.0143$ that survives a conservative significance test, yet at the F1-optimal threshold this prevents only six additional defaults out of 1,422 and cuts cost-weighted loss by under two percent. The explanatory half fails where the literature usually does not look: in the audited worked case, the generated narrative asserted that person_age ($-0.0092$), person_income ($-0.0569$), and loan_intent_VENTURE ($-0.0720$) raise risk even though the fused attributions supplied in the prompt recorded all three as lowering risk; it never mentioned loan_amnt ($+0.3297$), the strongest driver; and it introduced person_age, which was not in the supplied driver set. The paper ties this to measured properties of the attribution layer rather than to a one-off glitch: SHAP and LIME agree on which features matter (overlap@10 $= 0.80$) but not on their order ($\tau = 0.43$, $p = 0.18$), and the sign of the attribution for the ensemble's most sensitive input, person_income, flips near a coin flip across applicants (modal-sign share $0.53$). From this it concludes that constrained prompting is necessary but not sufficient, and that the three audit checks would have blocked the failure.

Load-bearing premise

The audit treats the sign of the fused SHAP-plus-LIME attribution as the ground truth the narrative must match; if that fused sign is itself an artifact of averaging two methods that disagree on ordering, the documented inversion shows the model disobeyed the prompt but not that it misrepresented the model's true risk direction.

Editorial extensions

If this is right

  • If the central claim is right, any deployment that treats prompt design as the safety mechanism is relying on a control with no enforcement, so generation-side verification becomes a mandatory component rather than an optional refinement.
  • The same audit checks are cheap enough to run on every decision: they require no model calls, no labels, and no human review, and they would have blocked the documented failure.
  • Explanation quality inherits attribution quality, so sign stability and inter-method rank agreement should be reported alongside importance rankings; without them, directional statements about an applicant are not supportable.
  • Because the ensemble's 0.0143 AUC gain translates into under two percent cost reduction at the deployed threshold, threshold selection against an explicit cost ratio matters more than further architecture search.
  • A narrative layer in a regulated setting should fall back to a deterministic template whenever the audit checks fail.

Reading between the lines

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

  • Editorial inference: If the fused SHAP-LIME sign is itself an artifact of averaging two methods that disagree on order, then the documented sign inversion may understate or misstate the model's true risk direction; a SHAP-only or LIME-only reference would disentangle the language model's defiance from the fusion step's ambiguity.
  • Editorial inference: The one documented case cannot support a failure rate, but the paper's own audit definition makes a rate study cheap: scoring several hundred narratives against the three checks would turn the existence proof into a systematic measurement and allow comparison across model scales and prompt designs.
  • Editorial inference: The same verify-after-generation pattern likely transfers to other regulated domains with generated explanations, such as medical or hiring decision letters, where a fluent but sign-inverted rationale would have similar consequences.
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 / 6 minor

Summary. This paper builds an end-to-end credit-risk system consisting of a multi-scale stacking ensemble (four gradient-boosting learners plus a residual MLP fused by a neural meta-learner trained on out-of-fold predictions) and an LLM-based explanation layer that feeds the top-four fused SHAP/LIME drivers to Phi-2 under a constrained prompt. The predictive claim is that the ensemble reaches test ROC-AUC 0.9539 and PR-AUC 0.9137, with a statistically significant but operationally small improvement over a tuned random forest. The explanatory claim is that, despite constrained prompting, the LLM-generated narrative in an audited case inverted the sign of three of four supplied drivers, omitted the dominant driver, and introduced a feature never given to it. The paper also reports trustworthiness diagnostics showing calibration and perturbation stability below the authors' own thresholds and concludes that constrained prompting is necessary but not sufficient, so grounding must be verified after generation.

Significance. If the narrative-failure finding is correct, it is an important negative result for a widely proposed architecture: input-side constraints on LLM-generated credit explanations do not guarantee grounded output, and cheap post-generation audits are necessary. The paper's strengths include leakage-controlled preprocessing, out-of-fold meta-features, conservative AUC significance testing operating under an independence assumption that disadvantages the authors' own claim, an operating-point cost analysis that qualifies the AUC gain, released reproduction scripts, and honest reporting of failures (miscalibration, perturbation instability, sign instability). The audit methodology (membership, direction, coverage) is simple and transferable. However, the strength of the central exhibit is reduced by an internal inconsistency between the prompt specification and the audit's sign-inversion count, as detailed below.

major comments (3)
  1. [Section X.B / Table V / Appendix A] The sign-inversion count is inflated because the prompt supplies signed impacts only for the top four drivers. In Table V the top-four set is {loan_amnt, person_home_ownership_RENT, person_home_ownership_OWN, loan_intent_VENTURE}; person_income (−0.0569) and person_age (−0.0092) are fifth and sixth and therefore were not among the signed entries in the prompt. Consequently, the narrative's statements about age and income are unsupported directional assertions, not inversions of supplied evidence. The only cleanly supported sign inversion is loan_intent_VENTURE (−0.0720 supplied as decreasing risk, narrative asserts increasing). The coverage failure (loan_amnt omitted) is also supported. The headline 'three of four checkable propositions invert the sign of the evidence' (Abstract, Section X.B, Section XIV) must be corrected, because it is material to the paper's central claim.
  2. [Abstract and Section XIV] The claim that the model 'introduced a feature never given to it' is contradicted by the prompt template in Appendix A, which includes the raw applicant profile with person_age. The model was given person_age; it was not given person_age in the signed driver list. Section X.B item 3 states this correctly ('not among the four drivers supplied'), but the abstract and conclusion overstate it. Please revise to avoid the inaccurate 'never given' phrasing.
  3. [Sections V.A, IX.C, X.A] The audit uses the fused SHAP+LIME sign as the ground truth that the narrative must match, but the paper's own measurements show this reference is fragile: SHAP and LIME agree on set membership (overlap@10 = 0.80) but not on order (tau = 0.43, p = 0.18), and the modal-sign share for person_income is 0.53 across instances. For loan_intent_VENTURE, the sign is relatively stable (79% modal-sign share in Fig. 8a), so the single clean inversion is robust. For person_income, a 'sign inversion' claim is of questionable meaning because the fusion reference itself is near-random in direction. The audit definition should explicitly state that the reference is the fused attribution, not a verified ground-truth direction, and the paper should temper the language accordingly.
minor comments (6)
  1. [References] References [10]–[13] are listed without author names; please complete them.
  2. [Section VI.B, Eq. (7)] The denominator of the Hanley–McNeil variance term appears as '1/n+n−'; this should be 1/(n_+ n_−) with n_+ and n_− the numbers of positive and negative cases.
  3. [Table V] Table V would be clearer if the 'Impact' column explicitly stated that these are fused min–max-normalized SHAP/LIME scores, not raw attribution values.
  4. [Section IX.C] Section IX.C reports Kendall tau and Spearman rho with p-values for n = 8; please state in the text that n is the number of features in the common top-10 set.
  5. [Section V.A] SHAP attributions are computed for 100 test instances and LIME for 50; please justify the different sample sizes.
  6. [Figure 8(b)] The x-axis label 'mean |Δp| under ± perturbation' should specify the perturbation strength (0.01) and units.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the predictive and narrative results are empirical comparisons against independent test data and externally supplied prompts; the audit is a check, not a derivation from its inputs.

full rationale

The paper's predictive claim (multi-scale stacking AUC 0.9539) is tested on a held-out public-data split against standard baselines, and the ensemble-versus-random-forest margin is assessed with a conservative significance test, so the performance result is not an artifact of fitting. The narrative-fidelity finding compares generated text to attribution values computed independently of the generation process; the audit criteria (membership, direction, coverage) are applied after generation and are not used to construct the prompt's driver set. No parameter is fitted to the audited output and then reported as a prediction; no self-citation is load-bearing, as the cited works are external XAI, LLM, and algorithmic references; and the SHAP/LIME stability measurements are made on held-out instances rather than derived from the conclusion. The self-declared calibration and stability thresholds are transparent acceptance criteria, not fitted inputs. One evidence-support issue is flagged but is not circularity: Appendix A states that signed impacts are provided only for the 'top 4' local drivers, while Table V audits signed impacts for person_income and person_age, which are fifth and sixth by magnitude; the cleanly supported sign inversion is loan_intent_VENTURE, and the omission of loan_amnt remains a coverage failure. This weakens the 'three of four' headline but does not reduce any derived claim to its input by construction. The paper is self-contained against external benchmarks and its limitations (single dataset, case-study audit, no calibration stage) are explicitly disclosed, further supporting a no-circularity verdict.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The central claims depend on a series of modeling choices and assumptions. The predictive result depends on hyperparameters and preprocessing choices that are set by hand or prior experiments, not fitted to the test set. The narrative-failure result depends on the assumption that the fused SHAP/LIME attribution is a valid reference, which the paper's own diagnostics weaken. No new physical entities are introduced.

free parameters (8)
  • F1-optimal threshold tau* = 0.362
    Selected on training data (Eq. 5) to set the operating point; used for all operating-point and cost comparisons.
  • Focal loss alpha, gamma = alpha=0.25, gamma=2.0
    Chosen constants for imbalance handling in neural base learners (Eq. 2).
  • Base learner hyperparameters = Table II values (depth 3-8, trees 300-500, eta 0.02-0.05)
    Set from prior experiments and deliberately spread across the depth-shrinkage plane, not tuned on the test set.
  • Meta-learner regularization = L2=1e-4, dropout=0.2, batch norm
    Chosen architecture and regularization for the 5->32->16->1 meta-learner.
  • Attribution sample sizes = SHAP 100 test instances, 200-sample background; LIME 50 instances
    Subsample sizes for computing SHAP and LIME; affect precision of all attribution-stability metrics.
  • Perturbation strength and repetitions = 0.01, 10 repetitions
    Parameters of the DPD perturbation-stability diagnostic.
  • Winsorization caps = age 100, emp_length 50
    Domain-imposed caps estimated on training data.
  • Imputation k = k=5
    Choice for KNN imputation of missing loan_int_rate and emp_length.
assumptions (4)
  • domain assumption SHAP permutation explainer and LIME local surrogates produce valid estimates of the ensemble's feature attributions.
    Used as the basis for both the driver list handed to the LLM and the audit ground truth (Section V.A). If these attributions are inaccurate, the sign-inversion finding is about faithfulness to the prompt, not to the model.
  • ad hoc to paper Averaging min-max normalized SHAP and LIME scores produces a meaningful fused driver set with interpretable signs.
    The fusion equation in Section V.A and the audit in Section X treat the fused sign as the evidence; Section IX.C shows the two methods disagree on ordering, which undermines this premise.
  • domain assumption The public Credit Risk Dataset is a sufficient testbed for the claimed behavior.
    All conclusions come from a single dataset with no temporal structure (Section III.A and XIII), so generalization across credit cycles and feature distributions is assumed, not shown.
  • domain assumption Deterministic greedy decoding makes the generated narrative reproducible.
    The audit assumes do_sample=False yields a stable output, which is reasonable for a fixed model checkpoint but depends on the exact model release and software versions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Accurate Ensembles, Fragile Narratives: Multi-Scale Stacking and a Fidelity Audit of LLM-Generated Explanations for Credit Risk." pith.science (2026). https://pith.science/paper/327WZ6NP

@misc{pith2026260808126,
  author       = {Pith},
  title        = {Pith review of: Accurate Ensembles, Fragile Narratives: Multi-Scale Stacking and a Fidelity Audit of LLM-Generated Explanations for Credit Risk},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/327WZ6NP}},
  note         = {Machine review of arXiv:2608.08126}
}
read the original abstract

Credit scoring increasingly relies on models whose decision logic cannot be read off their parameters, in tension with supervisory expectations that adverse decisions be explainable. A common proposal closes that gap with a language model: compute feature attributions, hand them to an LLM, and let it write the rationale. We build such a system end to end and test whether the second half of the promise holds. The predictive component is a multi-scale stacking ensemble fusing four differently regularised gradient-boosting learners with a residual network through a neural meta-learner trained on out-of-fold predictions. On a public 32,581-application credit dataset it reaches test ROC-AUC 0.9539 (95% CI [0.9462, 0.9616]) and PR-AUC 0.9137, beating the best single model by Delta-AUC = 0.0143 (p = 0.016 under a conservative independence assumption). Our central finding is asymmetric. The ranking gain is real but operationally small: at the F1-optimal threshold the ensemble avoids only six additional missed defaults out of 1,422 against a tuned random forest, cutting cost-weighted loss by under 2%. The narrative layer fails in a way prompt engineering alone does not fix. In an audited case the model named three factors as risk-increasing that the supplied attributions scored as risk-reducing, omitted the dominant driver, and introduced a feature never given to it. We trace this to properties we measure rather than assume: SHAP and LIME agree on which features matter (overlap@10 = 0.80) but not on their order (tau = 0.43, p = 0.18), and the attribution sign for the model's most sensitive input is near a coin flip across applicants (modal-sign share 0.53). Calibration (ECS = 0.117) and perturbation stability (DPD = 0.078) both fall short of our own thresholds. Constrained prompting is necessary but not sufficient: grounding must be verified after generation, not assumed.

Figures

Figures reproduced from arXiv: 2608.08126 by the authors.

Figure 1
Figure 1. End-to-end system. Five heterogeneous base learners are fused by a neural meta-learner trained on out-of-fold predictions; the resulting probability [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Out-of-fold construction of the meta-feature matrix. Within each of five stratified folds the base learners are refitted on the complementary four folds [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Test performance with uncertainty. (a) ROC-AUC with Hanley–McNeil [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 6
Figure 6. Figure 6: Per-base-learner SHAP importances, normalised by column max [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Global rankings under SHAP and LIME. Grey squares mark features [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: (a) Directional stability: share of the 100 explained instances on which a feature’s attribution takes its modal sign. Bars below 0.7 (red) indicate features whose risk direction routinely flips across applicants; a value near 0.5 means the direction is near-arbitrary …
Figure 9
Figure 9. Figure 9: Reliability diagram for the ensemble (n = 500, 10 bins, ECS = 0.1171). The model is over-confident below pˆ ≈ 0.3 and erratic above it. The upper bins are sparsely populated, so the extreme excursions should be read as high-variance rather than as evidence of systemati…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 30 canonical work pages

  1. [1]

    European Union regulations on algorith- mic decision-making and a ‘right to explanation’,

    B. Goodman and S. Flaxman, “European Union regulations on algorith- mic decision-making and a ‘right to explanation’,”AI Magazine, vol. 38, no. 3, pp. 50–57, 2017

    work page 2017

  2. [2]

    Benchmark- ing state-of-the-art classification algorithms for credit scoring: An update of research,

    S. Lessmann, B. Baesens, H.-V . Seow, and L. C. Thomas, “Benchmark- ing state-of-the-art classification algorithms for credit scoring: An update of research,”European Journal of Operational Research, vol. 247, no. 1, pp. 124–136, 2015

    work page 2015

  3. [3]

    Statistical and machine learning models in credit scoring: A systematic literature survey,

    X. Dastile, T. Celik, and M. Potsane, “Statistical and machine learning models in credit scoring: A systematic literature survey,”Applied Soft Computing, vol. 91, 106263, 2020

    work page 2020

  4. [4]

    A unified approach to interpreting model predictions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” inAdvances in Neural Information Processing Systems, vol. 30, 2017, pp. 4765–4774

    2017

  5. [5]

    ‘Why should I trust you?’ Ex- plaining the predictions of any classifier,

    M. T. Ribeiro, S. Singh, and C. Guestrin, “‘Why should I trust you?’ Ex- plaining the predictions of any classifier,” inProc. 22nd ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining, 2016, pp. 1135–1144

    work page 2016

  6. [6]

    Explainable machine learning in credit risk management,

    N. Bussmann, P. Giudici, D. Marinelli, and J. Papenbrock, “Explainable machine learning in credit risk management,”Computational Economics, vol. 57, pp. 203–216, 2021

    work page 2021

  7. [7]

    LLMs for XAI: Future directions for converting formal explanations to natural language,

    A. Zytek, S. Pidò, and K. Veeramachaneni, “LLMs for XAI: Future directions for converting formal explanations to natural language,”arXiv preprintarXiv:2405.13036, 2024

    arXiv 2024

  8. [8]

    XAI for all: Can large language models simplify explainable AI?,

    P. Mavrepis, G. Makridis, G. Fatouros, V . Koukos, M. M. Separdani, and D. Kyriazis, “XAI for all: Can large language models simplify explainable AI?,”arXiv preprintarXiv:2401.13110, 2024

    arXiv 2024

Show all 45 references
  1. [9]

    Are large language models post hoc explainers?,

    N. Kroeger, D. Ley, S. Krishna, C. Agarwal, and H. Lakkaraju, “Are large language models post hoc explainers?,”arXiv preprint arXiv:2310.05797, 2023

    2023 arXiv

  2. [10]

    From XAI to stories: A factorial study of LLM-generated explanation quality,

    “From XAI to stories: A factorial study of LLM-generated explanation quality,”arXiv preprintarXiv:2601.02224, 2026

    2026

  3. [11]

    Could large language models work as post-hoc explainability tools in credit risk models?,

    “Could large language models work as post-hoc explainability tools in credit risk models?,”arXiv preprintarXiv:2602.18895, 2026

    2026 arXiv

  4. [12]

    A two-stage LLM framework for accessible and verified XAI explana- tions,

    “A two-stage LLM framework for accessible and verified XAI explana- tions,”arXiv preprintarXiv:2604.12543, 2026

    2026 arXiv

  5. [13]

    Interpreting LLMs as credit risk classifiers: Do their feature explana- tions align with classical ML?,

    “Interpreting LLMs as credit risk classifiers: Do their feature explana- tions align with classical ML?,”arXiv preprintarXiv:2510.25701, 2025

    2025

  6. [14]

    Survey of hallucination in natural language generation,

    Z. Jiet al., “Survey of hallucination in natural language generation,” ACM Computing Surveys, vol. 55, no. 12, pp. 1–38, 2023

    2023

  7. [15]

    Towards faithfully interpretable NLP systems: How should we define and evaluate faithfulness?,

    A. Jacovi and Y . Goldberg, “Towards faithfully interpretable NLP systems: How should we define and evaluate faithfulness?,” inProc. 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 4198–4205

    2020

  8. [16]

    Fooling LIME and SHAP: Adversarial attacks on post hoc explanation methods,

    D. Slack, S. Hilgard, E. Jia, S. Singh, and H. Lakkaraju, “Fooling LIME and SHAP: Adversarial attacks on post hoc explanation methods,” in Proc. AAAI/ACM Conf. AI, Ethics, and Society, 2020, pp. 180–186

    2020

  9. [17]

    On the robustness of interpretabil- ity methods,

    D. Alvarez-Melis and T. S. Jaakkola, “On the robustness of interpretabil- ity methods,”arXiv preprintarXiv:1806.08049, 2018

    2018 arXiv

  10. [18]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,

    C. Rudin, “Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,”Nature Machine Intelligence, vol. 1, pp. 206–215, 2019

    2019

  11. [19]

    Why do tree-based models still outperform deep learning on typical tabular data?,

    L. Grinsztajn, E. Oyallon, and G. Varoquaux, “Why do tree-based models still outperform deep learning on typical tabular data?,” in Advances in Neural Information Processing Systems, Datasets and Benchmarks Track, vol. 35, 2022

    2022

  12. [20]

    Tabular data: Deep learning is not all you need,

    R. Shwartz-Ziv and A. Armon, “Tabular data: Deep learning is not all you need,”Information Fusion, vol. 81, pp. 84–90, 2022

    2022

  13. [21]

    XGBoost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” inProc. 22nd ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining, 2016, pp. 785–794

    2016

  14. [22]

    LightGBM: A highly efficient gradient boosting decision tree,

    G. Keet al., “LightGBM: A highly efficient gradient boosting decision tree,” inAdvances in Neural Information Processing Systems, vol. 30, 2017, pp. 3146–3154

    2017

  15. [23]

    CatBoost: Unbiased boosting with categorical features,

    L. Prokhorenkova, G. Gusev, A. V orobev, A. V . Dorogush, and A. Gulin, “CatBoost: Unbiased boosting with categorical features,” inAdvances in Neural Information Processing Systems, vol. 31, 2018

    2018

  16. [24]

    Random forests,

    L. Breiman, “Random forests,”Machine Learning, vol. 45, no. 1, pp. 5– 32, 2001

    2001

  17. [25]

    Stacked generalization,

    D. H. Wolpert, “Stacked generalization,”Neural Networks, vol. 5, no. 2, pp. 241–259, 1992

    1992

  18. [26]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2016, pp. 770–778

    2016

  19. [27]

    TabNet: Attentive interpretable tabular learning,

    S. Ö. Arik and T. Pfister, “TabNet: Attentive interpretable tabular learning,” inProc. AAAI Conf. Artificial Intelligence, vol. 35, no. 8, 2021, pp. 6679–6687

    2021

  20. [28]

    Deep & cross network for ad click predictions,

    R. Wang, B. Fu, G. Fu, and M. Wang, “Deep & cross network for ad click predictions,” inProc. ADKDD’17, 2017, pp. 1–7

    2017

  21. [29]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal loss for dense object detection,” inProc. IEEE Int. Conf. Computer Vision, 2017, pp. 2980–2988

    2017

  22. [30]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inProc. Int. Conf. Learning Representations, 2015

    2015

  23. [31]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” inProc. Int. Conf. Machine Learning, 2015, pp. 448–456

    2015

  24. [32]

    Dropout: A simple way to prevent neural networks from overfit- ting,

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut- dinov, “Dropout: A simple way to prevent neural networks from overfit- ting,”Journal of Machine Learning Research, vol. 15, pp. 1929–1958, 2014

    1929

  25. [33]

    Searching for activation functions,

    P. Ramachandran, B. Zoph, and Q. V . Le, “Searching for activation functions,”arXiv preprintarXiv:1710.05941, 2017

    2017 arXiv

  26. [34]

    A value forn-person games,

    L. S. Shapley, “A value forn-person games,” inContributions to the Theory of Games II, Princeton Univ. Press, 1953, pp. 307–317

    1953

  27. [35]

    Explaining prediction models and individual predictions with feature contributions,

    E. Štrumbelj and I. Kononenko, “Explaining prediction models and individual predictions with feature contributions,”Knowledge and In- formation Systems, vol. 41, pp. 647–665, 2014

    2014

  28. [36]

    A survey of methods for explaining black box models,

    R. Guidotti, A. Monreale, S. Ruggieri, F. Turini, F. Giannotti, and D. Pedreschi, “A survey of methods for explaining black box models,”ACM Computing Surveys, vol. 51, no. 5, pp. 1–42, 2018

    2018

  29. [37]

    Explainable Artificial Intelligence (XAI): Con- cepts, taxonomies, opportunities and challenges toward responsible AI,

    A. B. Arrietaet al., “Explainable Artificial Intelligence (XAI): Con- cepts, taxonomies, opportunities and challenges toward responsible AI,” Information Fusion, vol. 58, pp. 82–115, 2020

    2020

  30. [38]

    Counterfactual explanations without opening the black box: Automated decisions and the GDPR,

    S. Wachter, B. Mittelstadt, and C. Russell, “Counterfactual explanations without opening the black box: Automated decisions and the GDPR,” Harvard Journal of Law & Technology, vol. 31, no. 2, pp. 841–887, 2017

    2017

  31. [39]

    On calibration of modern neural networks,

    C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” inProc. Int. Conf. Machine Learning, 2017, pp. 1321–1330

    2017

  32. [40]

    Predicting good probabilities with supervised learning,

    A. Niculescu-Mizil and R. Caruana, “Predicting good probabilities with supervised learning,” inProc. Int. Conf. Machine Learning, 2005, pp. 625–632

    2005

  33. [41]

    Equality of opportunity in super- vised learning,

    M. Hardt, E. Price, and N. Srebro, “Equality of opportunity in super- vised learning,” inAdvances in Neural Information Processing Systems, vol. 29, 2016

    2016

  34. [42]

    The meaning and use of the area under a receiver operating characteristic (ROC) curve,

    J. A. Hanley and B. J. McNeil, “The meaning and use of the area under a receiver operating characteristic (ROC) curve,”Radiology, vol. 143, no. 1, pp. 29–36, 1982

    1982

  35. [43]

    Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach,

    E. R. DeLong, D. M. DeLong, and D. L. Clarke-Pearson, “Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach,”Biometrics, vol. 44, no. 3, pp. 837– 845, 1988

    1988

  36. [44]

    Probable inference, the law of succession, and statistical inference,

    E. B. Wilson, “Probable inference, the law of succession, and statistical inference,”Journal of the American Statistical Association, vol. 22, no. 158, pp. 209–212, 1927

    1927

  37. [45]

    Phi-2: The surprising power of small language models,

    Microsoft Research, “Phi-2: The surprising power of small language models,” 2023. [Online]. Available: https://www.microsoft.com/en-us/ research/blog/phi-2-the-surprising-power-of-small-language-models/

    2023

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.