A unified framework for global and local interpretability using adaptive derivative-ordered random explanation
Lemen Chao, Ming Lei, Anran Fang
cs.LG · cs.AI
Quick verdict
ADORE unifies global/local XAI via derivatives + randomized SVD; efficiency gains look real, but flawed adaptivity logic and no significance tests temper it.
Summary
The paper introduces ADORE (Adaptive Derivative-Ordered Random Explanation), a post-hoc, model-agnostic interpretability method that builds a derivative matrix over samples and features using central finite differences, adaptively choosing first- or second-order approximations per sample via a curvature-based complexity measure, then compressing the matrix with randomized SVD and sparsity thresholding. Global feature importance, per-sample signed contributions, sample leverage scores, and feature importance scores are all derived from this single matrix. Experiments cover three modalities: California Housing with Random Forest (ADORE ties PFI at 98.91% fidelity at k=2 while running in 9.80 s vs SHAP's 19.03 s and PFI's 23.36 s, with 0.39 MB peak memory), IMDB sentiment with fine-tuned BERT (ADORE: 29% fidelity at k=10, 14.09 s vs SHAP 91.61 s), and CIFAR-10 with ResNet-50 (qualitative VGG16-feature-map visualizations only). Key limitations: the adaptivity rule in Section 3.2.2 appears inverted (second order selected when the curvature measure is below threshold); Eq. 10 computes only diagonal second differences, so cross-feature interactions, the paper's headline motivation, are not actually quantified; Eq. 15-18 contain formula inconsistencies; no statistical uncertainty, seeds, or significance tests accompany the comparisons; hyperparameter values (threshold tau, rank k) are unreported; and the image pipeline relies on VGG16 hidden layers, conflicting with the model-agnostic claim. Reproducibility is partially supported by a released Python package.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single derivative matrix built by adaptive finite differences provides both global feature importance and local sample contributions in one framework.
- Per-sample adaptive switching between first- and second-order derivatives handles both linear and nonlinear model regions.
- Randomized SVD plus sparsity thresholding yields scalability, with total complexity O(m·n·Cf + m·n·k).
- ADORE matches or exceeds LIME/SHAP/PFI/IG on fidelity and efficiency across tabular and text tasks, and is model-agnostic over tabular, text and image inputs.
Strengths
Unified global+local attribution from one object
The m×n derivative matrix yields signed per-sample contributions (Eq. 15-16), global feature importance (Eq. 18) and sample leverage scores (Eq. 17), directly answering the fragmentation critique of Section 2.
Strong efficiency profile on tabular and text
Table 3: ADORE 9.80 s / 0.39 MB vs SHAP 19.03 s / 3.74 MB and PFI 23.36 s / 367.88 MB; Table 4: 14.09 s vs SHAP 91.61 s, IG 65.00 s, LIME 300.51 s — fastest on text, and by far the lowest tabular memory.
Transparent cost accounting
Fidelity parity with PFI at a fraction of the cost
Artifact released
Weaknesses
Adaptive rule appears inverted
Headline interaction claim not supported by the math
No statistical uncertainty
Image modality is qualitative only and strains the model-agnostic claim
Reproducibility and editorial gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
ADORE: Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the ADORE algorithm?
ADORE (Adaptive Derivative-Ordered Random Explanation) is a post-hoc interpretability method for black-box machine learning models. It perturbs each input feature, estimates first- and second-order derivatives by finite differences, stacks them into a sample-by-feature derivative matrix, compresses that matrix with randomized SVD, and derives global feature importance, per-sample signed contributions, and influential samples from the single matrix.
How does ADORE differ from SHAP and LIME?
LIME fits local linear surrogates and SHAP averages marginal contributions over feature subsets; both can be slow on high-dimensional data. ADORE instead uses only about 2n+1 model evaluations per sample for derivative estimates and compresses via randomized SVD. In the paper's runs it was faster than SHAP (9.80 s vs 19.03 s on tabular data; 14.09 s vs 91.61 s on text), though fidelity differences were small and reported without significance tests.
Is ADORE's code available?
Yes. The authors released a Python package called adore (version 0.1.2) on PyPI, linked from the paper. However, some experimental settings used in the paper — the derivative-order threshold tau and the SVD rank k — are not reported numerically, so exact replication of the result tables may require additional work.
What datasets and models were used to evaluate ADORE?
Three tasks: California Housing price regression with a Random Forest (RMSE 0.49, R-squared 0.82), IMDB movie-review sentiment with a fine-tuned bert-base-uncased checkpoint (91% accuracy), and CIFAR-10 image classification with a fine-tuned ResNet-50 (94.7% accuracy). The image-modality results are qualitative visualizations only.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2-5 are referenced but not analyzable from the extracted text; image-modality and visualization judgments rest on the prose alone.
- Several equations (notably Eq. 15 and Eq. 17) appear truncated or corrupted by extraction; formula criticisms reflect the text as provided.
- This is an arXiv preprint (2609.17171); peer-review status is unknown.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.