Paper: arXiv 2309.13246

Authors: Dangxing Chen

Abstract

In recent years, explainable machine learning methods have been very successful. Despite their success, most explainable machine learning methods are applied to black-box models without any domain knowledge. By incorporating domain knowledge, science-informed machine learning models have demonstrated better generalization and interpretation. But do we obtain consistent scientific explanations if we apply explainable machine learning methods to science-informed machine learning models? This question is addressed in the context of monotonic models that exhibit three different types of monotonicity. To demonstrate monotonicity, we propose three axioms. Accordingly, this study shows that when only individual monotonicity is involved, the baseline Shapley value provides good explanations; however, when strong pairwise monotonicity is involved, the Integrated gradients method provides reasonable explanations on average.

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 introduces formal axioms and theoretical comparisons between Integrated Gradients and Baseline Shapley, relying on derivations and proofs that demonstrate high mathematical density. However, the provided excerpt and summary focus on theoretical properties without any discussion of backtesting, specific datasets, code implementations, or statistical metrics, indicating low empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal<br>How do XAI methods perform<br>on monotonic science-informed models?"] --> B{"Methodology"}
  B --> C["Define 3 Monotonicity Axioms"]
  C --> D["Data: Synthetic & Financial Datasets"]
  D --> E["Models: Monotonicity-enforced<br>ML (e.g., Neural Networks)"]
  E --> F["Process: Apply & Compare<br>Shapley Values vs. Integrated Gradients"]
  F --> G{"Key Findings"}
  G --> H["Individual Monotonicity:<br>Shapley Values perform well"]
  G --> I["Strong Pairwise Monotonicity:<br>Integrated Gradients perform better"]

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 three new axioms for average preservation of three types of monotonicity and analyzes which attribution methods satisfy them.

Source sentence
We propose three new axioms concerning the average preservation of three types of monotonicity.

Assumes 4

  • A zero baseline (x′ = 0) is used for attribution methods. model
    Source sentence
    For simplicity, we consider the zero baseline as x′ = 0.
  • Missing variables are removed and past dues greater than four times are truncated in the credit dataset. data
    Source sentence
    For simplicity, data with missing variables are removed. Past dues greater than four times are truncated.
  • The dataset is randomly partitioned into 75% training and 25% test sets. data
    Source sentence
    The dataset is randomly partitioned into 75% training and 25% test sets.
  • Features x1‑x3 are grouped together and monotonicity is only imposed on these features. model
    Source sentence
    All functions are approximated by neural networks with one hidden layer and two neurons. For

Data

Recognised sources: Kaggle

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • Baseline Shapley (BShap) provides satisfactory explanations when only individual monotonicity is involved. theoretical
    Source sentence
    BShap provides satisfactory explanations when only individual monotonicity is involved
  • Integrated Gradients (IG) provides reasonable explanations when strong pairwise monotonicity is involved. theoretical
    Source sentence
    IG provides reasonable explanations when strong pairwise monotonicity is involved
  • Both the black‑box FCNN and the transparent monotonic neural additive model achieve an AUC of approximately 80%. performance
    Source sentence
    Both models have an AUC of approximately 80%.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 22 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 3 papers in the archive (2 by independent authors)

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.