Paper: arXiv 2408.05701

Authors: Dangxing Chen, Jingfeng Chen, Weicheng Ye

Abstract

Explainable machine learning methods have been accompanied by substantial development. Despite their success, the existing approaches focus more on the general framework with no prior domain expertise. High-stakes financial sectors have extensive domain knowledge of the features. Hence, it is expected that explanations of models will be consistent with domain knowledge to ensure conceptual soundness. In this work, we study the group structures of features that are naturally formed in the financial dataset. Our study shows the importance of considering group structures that conform to the regulations. When group structures are present, direct applications of explainable machine learning methods, such as Shapley values and Integrated Gradients, may not provide consistent explanations; alternatively, group versions of the Shapley value can provide consistent explanations. We contain detailed examples to concentrate on the practical perspective of our framework.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 2.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper is mathematically dense, featuring formal axioms, theorems, and proofs related to Shapley values and group structures. However, it lacks any empirical validation such as backtests, datasets, or code implementation, focusing instead on theoretical frameworks and conceptual examples.

Research Flowchart

  flowchart TD
  A["Research Goal: Reconcile XAI with financial domain knowledge?"] --> B{"Methodology: Analyze Financial Dataset"}
  B --> C["Key Finding: Group structures exist & align with regulations"]
  C --> D["Problem: Standard XAI<br>Shapley/IG inconsistent"]
  D --> E["Solution: Group-based XAI<br>Group-Shapley"]
  E --> F["Outcome: Consistent &<br>domain-conformant explanations"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes group axioms for attribution methods and introduces a group version of the Shapley value (GShap) that preserves these axioms.

Source sentence
(1) we propose group axioms for attribution methods with a rigorous theoretical framework.

Assumes 3

  • The dataset D has n samples and m features, and the model f is differentiable almost everywhere. data
    Source sentence
    For problem setup, assume we have D ×Y, where D is the dataset with
  • Zero baselines are used for attribution calculations. model
    Source sentence
    For simplicity, we consider the zero baselines as x′ = 0.
  • Group structures are present in the feature set and should be respected by attribution methods. other
    Source sentence
    When group structures are present, direct applications of explainable machine learning methods, such as Shapley values and Integrated Gradients, may not provide consistent explanations; alternatively, group versions of the Shapley value can provide consistent explanations.

Data

  • credit scoring dataset · Kaggle fundamentals public
    Source sentence
    We examine the significance of group structure using a widely-used credit scoring dataset from Kaggle, publicly available at 3.

Costs

Not applicable The paper reports no trading or execution result.

Finds 2

  • GShap preserves all knowledge-inspired axioms (ASI, DIM, GASI, GLFI, GDIM, GSPM) whereas BShap and IG do not. theoretical
    Source sentence
    As a result of the preservation of axioms when group structures are present, GShap is preferred over BShap and IG.
  • BShap and IG can produce absurd attributions for past‑due features, reducing attribution when past‑due payments increase. negative
    Source sentence
    In the content of Example 4.13, according to BShap and IG, for a candidate with more past‑due payments, the overall past‑due attributions could be significantly reduced, which is absurd.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 7 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (140)

  1. Statistical classification methods in consumer credit scoring: a review (1997)
  2. A credit scoring model for personal loans (1989)
  3. Benchmarking state-of-the-art classification algorithms for credit scoring: An update of research (2015)
  4. Benchmarking state-of-the-art classification algorithms for credit scoring (2003)
  5. The comparisons of data mining techniques for the predictive accuracy of probability of default of credit card clients (2009)
  6. Multiple classifier architectures and their application to credit risk assessment (2011)
  7. An interpretable model with globally consistent explanations for credit risk (2018) — link
  8. Boolean decision rules via column generation (2018) — link
and 132 more
  1. ViCE: visual counterfactual explanations for machine learning models (2020)
  2. Significance tests for neural networks (2020)
  3. Selectivenet: A deep neural network with an integrated reject option (2019)
  4. Selective classification for deep neural networks (2017) — link
  5. Recent developments in consumer credit risk assessment (2007)
  6. The MNIST database of handwritten digits (1999)
  7. Deep learning (2015)
  8. Learning multiple layers of features from tiny images (2009)
  9. Introduction to reinforcement learning (1998)
  10. Pattern Recognition Letters (2005)
  11. On Optimal Recognition Error and Reject Tradeoff (1970)
  12. SelectiveNet: A Deep neural network with an integrated reject option (2019)
  13. Abstaining classification when error costs are unequal and unknown (2018) — link
  14. On optimal recognition error and reject Tradeoff (1970)
  15. A method for improving classification reliability of multilayer perceptrons (1995)
  16. On optimal reject rules and ROC curves (2005)
  17. An optimal reject rule for binary classifiers (2000)
  18. Bootstrap methods for reject rules of Fisher LDA (2006)
  19. An empirical comparison of ideal and empirical ROC-based reject rules (2007)
  20. Class noise vs. attribute noise: A quantitative study of their impacts (2004)
  21. Theoretical and experimental analysis of a Two-Stage system for classification (2002)
  22. Improving the classification accuracy of noisy dataset by effective data preprocessing (2018)
  23. Spectrally-normalized margin bounds for neural networks (2017)
  24. Multi-objective optimization for self-adjusting weighted gradient in machine learning tasks (2015) — link
  25. UCI machine learning repository (2017)
  26. Deep learning (2016)
  27. Show, attend and tell: Neural image caption generation with visual attention (2016) — link
  28. Very deep convolutional networks for large-scale image recognition (2015)
  29. Going deeper with convolutions (2015)
  30. Deep residual learning for image recognition (2016)
  31. Deep Reinforcement Learning: A brief survey (2017)
  32. Training deep neural-networks using a noise adaptation layer (2017)
  33. Making deep neural networks robust to label noise: A loss correction approach (2017)
  34. Understanding deep learning requires rethinking generalization (2017)
  35. Building high-level features using large scale unsupervised learning (2012)
  36. Unsupervised visual representation learning by context prediction (2015)
  37. The curious robot: Learning visual representations via physical interactions (2016)
  38. Colorful image colorization (2016)
  39. Webly supervised learning of convolutional networks (2015)
  40. Learning From noisy large-scale datasets With minimal supervision (2017)
  41. On the foundations of noise-free selective classification (2010)
  42. Selective classification for deep neural networks (2017) — link
  43. Support vector machines with embedded reject option (2002)
  44. The nearest neighbor classification rule with a reject option (1970)
  45. Boosting with abstention (2016)
  46. Reject option with multiple thresholds (2000)
  47. Neural networks for pattern recognition (1995)
  48. The multilayer perceptron as an approximation to a Bayes optimal discriminant function (1990)
  49. Methods for interpreting and understanding deep neural networks (2018)
  50. Sensitivity of feedforward neural networks to Weight Errors (1990)
  51. Sensitivity based Neural Networks Explanations (2018) — link
  52. Intelligible Models for Health-Care: Predicting Pneumonia Risk and Hospital 30-day Readmission (2015)
  53. Good Explanation for Algorithmic Transparency (2019)
  54. Explainable Artificial Intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI (2019) — link
  55. Learning Algorithm in Two-Stage Selective Prediction (2022)
  56. Advances in financial machine learning (2018)
  57. Approximation by superpositions of a sigmoidal function (1989)
  58. Approximation capabilities of multilayer feedforward networks (1991)
  59. Neural networks: a comprehensive foundation by Simon Haykin, Macmillan, 1994, ISBN 0-02-352781-7. (1999)
  60. Fundamentals of artificial neural networks (1995)
  61. Neural networks for optimal approximation of smooth and analytic functions (1996)
  62. The eu general data protection regulation (gdpr) (2017)
  63. Machine learning for credit scoring: Improving logistic regression with non-linear decision-tree effects (2022)
  64. Credit granting: A comparative analysis of classification procedures (1987)
  65. Construction of a k-nearest-neighbour credit-scoring system (1997)
  66. Credit scoring using neural and evolutionary techniques (2000)
  67. Subagging for credit scoring models (2010)
  68. Using Random Forest for credit risk models (2018)
  69. Certified monotonic neural networks (2020)
  70. Sparse convolutional neural networks (2015)
  71. Deep sparse rectifier neural networks (2011)
  72. Training sparse neural networks (2017)
  73. A Biometrics invited paper. The analysis and selection of variables in linear regression (1976)
  74. Neural additive models: Interpretable machine learning with neural nets (2021)
  75. Enhancing explainability of neural networks through architecture constraints (2020)
  76. GAMI-Net: An explainable neural network based on generalized additive models with structured interactions (2021)
  77. Unwrapping the black box of deep ReLU networks: interpretability, diagnostics, and simplification (2020) — link
  78. A unified approach to interpreting model predictions (2017)
  79. " Why should i trust you?" Explaining the predictions of any classifier (2016)
  80. Fast and flexible monotonic functions with ensembles of lattices (2016)
  81. Deep lattice networks and partial monotonic functions (2017)
  82. Deep residual learning for image recognition (2016)
  83. Language models are unsupervised multitask learners (2019)
  84. Designing Inherently Interpretable Machine Learning Models (2021) — link
  85. Pitfalls to avoid when interpreting machine learning models (2020)
  86. Problems with Shapley-value-based explanations as feature importance measures (2020)
  87. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead (2019)
  88. Fooling lime and shap: Adversarial attacks on post hoc explanation methods (2020)
  89. Re: Request for Information and Comment on Financial Institutions’ Use of Artificial Intelligence, including Machine Learning (Docket No. CFPB-2021-0004) (2021)
  90. Seven challenges for harmonizing explainability requirements (2021) — link
  91. AI regulation is coming (2021) — [link](https://hbr.org/2021/09/ai-regulation- is- coming.)
  92. Model risk management (2021) — link
  93. Fair lending (2010) — link
  94. Multicriteria interpretability driven deep learning (2022)
  95. The many Shapley values for model explanation (2020)
  96. A rigorous study of integrated gradients method and extensions to internal neuron attributions (2022)
  97. Axiomatic attribution for deep networks (2017)
  98. Shape constraints for set functions (2019)
  99. Monotonic Neural Additive Models: Pursuing Regulated Machine Learning Models for Credit Scoring (2022)
  100. Three methods to share joint costs or surplus (1999)
  101. Anchors: High-precision model-agnostic explanations (2018)
  102. Physics-informed machine learning (2021)
  103. Hamiltonian neural networks (2019)
  104. How to address monotonicity for model risk management? — → 2305.00799
  105. Integrated Gradients is a Nonlinear Generalization of the Industry Standard Approach to Variable Attribution for Credit Risk Models (2022)
  106. Explainable machine learning in credit risk management (2021)
  107. Multidimensional shape constraints (2020)
  108. Applications of Integrated Gradients in Credit Risk Modeling (2022)
  109. A two-step Shapley value for cooperative games with coalition structures (2009)
  110. Values of games with a priori unions (1977)
  111. Deontological ethics by monotonicity shape constraints (2020)
  112. The nature, importance, and difficulty of machine ethics (2006)
  113. Automation, algorithms, and politics| when the algorithm itself is a racist: Diagnosing ethical harm in the basic components of software (2016)
  114. The Shapley-solidarity value for games with a coalition structure (2013)
  115. A new axiomatization of the Shapley–solidarity value for games with a coalition structure (2018)
  116. The Shapley Value in Machine Learning (2022) — link
  117. Financial ratios, discriminant analysis and the prediction of corporate bankruptcy (1968)
  118. Harnessing prior knowledge for explainable machine learning: An overview (2023)
  119. In what ways are deep neural networks invariant and how should we measure this? (2022)
  120. Constrained monotonic neural networks (2023)
  121. Learning with Group Invariant Features: A Kernel Perspective. (2015)
  122. Active learning for imbalanced data under cold start (2021)
  123. Attribution Methods in Asset Pricing: Do They Account for Risk? — → 2407.08953
  124. The Shapley value decomposition of optimal portfolios (2021)
  125. Scilab — link
  126. Scilab — link
  127. The Computational Geometry Algorithms Library — link
  128. The Computational Geometry Algorithms Library — link
  129. The Parmap library — link
  130. The Parmap library — link
  131. BabyJ : from Object Based to Class Based Programming via Types (2003)
  132. tag-sound

2 of 140 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.