Assessing Alignment and Stability of Feature Importance Explanations via Weight of Evidence
Eddie Conti, Claudio Daka, Álvaro Parafita, Antonio L. Alfeo, Axel Brando, Mario G. C. A. Cimino
cs.LG · cs.AI
Quick verdict
Reframes FIM evaluation as Weight of Evidence hypothesis testing; neat variance-to-WoE theory, but evidence is small-scale tabular with strong assumptions.
Summary
The paper reframes the evaluation of feature importance methods (FIMs) such as LIME and SHAP within a Weight of Evidence (WoE) hypothesis-testing framework. Each run of a stochastic FIM yields an attribution vector, converted into a binary 'relevant feature set' hypothesis via top-k or adaptive (tau-fraction of attribution mass) selection. WoE then quantifies how strongly the N-run evidence supports a reference hypothesis h* — drawn from domain knowledge, ground truth, or the FIM's own mean attribution — against its complement, with a uniform prior over 2^d relevance assignments. Theoretically, the authors prove (via Chebyshev-inequality arguments) that low attribution variance implies stability of the selected feature set and that WoE diverges to +infinity as variance vanishes (Theorem 1). Experiments on Titanic (random forest) show LIME yields 27 +infinity and 23 finite positive WoE values against a sex+pclass reference, while deterministic SHAP gives 23 +infinity and 27 -infinity, exposing local-global discrepancy. A synthetic make_classification study (3 informative + 7 redundant features, varying noise lambda) shows WoE can localize misalignment along the data→model→FIM pipeline and diagnose LIME kernel-width unfaithfulness at lambda=0.65. Four tabular datasets confirm SHAP's higher internal consistency. Limitations: exact-set-match posterior, uniform prior, mutual-independence assumptions, and deferred preprocessing details.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 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
- Embedding FIMs in a WoE hypothesis-testing framework enables principled, contrastive evaluation of alignment (domain knowledge/ground truth), faithfulness, and stability within one formalism.
- Low attribution variability implies the WoE computed against a mean-attribution reference hypothesis diverges to +∞ with probability approaching 1 (Theorem 1).
- WoE exposes local–global discrepancy on Titanic and localizes misalignment along the data→model→FIM pipeline, including LIME kernel-width unfaithfulness at λ=0.65.
- SHAP produces more internally consistent (less variable) explanations than LIME across four tabular datasets.
Strengths
Novel evaluation framing with clear positioning
Section 2 explicitly contrasts treating an explanation as a testable hypothesis against prior WoE uses that generate or structure explanations (Melis et al. [21], Alshehri et al. [2], Le et al. [16], Parola et al. [24]); the same WoE machinery then yields three diagnostics — alignment, faithfulness, stability — across Sections 5.1–5.3.
Complete theory-to-experiment chain
Section 4 builds Theorem 1 (WoE → +∞ as variance → 0) from elementary tools (union bound, Chebyshev on partial sums in Proposition 1 with margin δ of eq. 7, pairwise Lemma 2 and Corollary 1, binomial tail Lemma 3), and Section 5.4 validates it both with LIME num_samples sweeps (100–5000; RF/NN/DT/SVM on Diabetes) and a synthetic variance-controlled FIM (σ=(1/l)·0.1).
Actionable diagnostic in Section 5.2
Candid limitations section
Weaknesses
Logically loose alternative-hypothesis accounting
±∞ degeneracy dominates headline results
Strong assumptions behind Theorem 1
Small scale and missing baselines
Claims without statistical tests
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper reframes evaluation of feature importance methods (FIMs) — with LIME and SHAP as case studies — as hypothesis testing under Weight of Evidence (WoE). Each of N runs on a fixed instance yields an attribution vector binarized into a relevant-feature-set hypothesis via top-k selection (k=3 in ground-truth settings) or adaptive selection (τ=0.8 of normalized attribution mass). WoE is the posterior minus prior log-odds, with a uniform prior over 2^d relevance assignments (p(h)=2^−|Rx|). Depending on the reference hypothesis h*, one mode tests alignment with domain knowledge (Titanic), alignment/faithfulness with ground truth and model ranking (synthetic data), or stability against the FIM's own mean attribution. Central claims: the framework is principled, contrastive, and flexible; Theorem 1 states WoE → +∞ as attribution variance → 0 with probability → 1; and experiments expose local–global discrepancy, localize misalignment along the data→model→FIM pipeline, and quantify internal consistency.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 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 Weight of Evidence (WoE) in this paper?
WoE is the log-odds update that observed evidence provides for a hypothesis versus an alternative: log[p(e|h)/p(e|h')]. Here the 'evidence' is N repeated runs of a feature importance method (like LIME or SHAP) on one data instance, and each run is converted into a binary hypothesis about which features are relevant.
What does it mean when WoE is +∞ or −∞?
It means every one of the N runs reproduced the reference hypothesis (+∞, estimated posterior probability 1) or none did (−∞, posterior probability 0), because log-odds diverge at probability boundaries. Many SHAP results in the paper are divergent since SHAP is nearly deterministic.
Which methods and datasets were used?
LIME and SHAP on tabular data: Titanic (random forest, with 'sex' and 'passenger class' as the domain-knowledge reference), a synthetic scikit-learn classification set (3 informative plus 7 redundant features with varying noise), and the Diabetes, Heart Disease, Mobile-phone-pricing, and Churn datasets for stability analysis.
Does the paper provide code?
No code or repository is mentioned. The authors state that dataset, preprocessing, and hyperparameter details 'will be left for the final version', so exact reproduction currently requires reimplementation from the paper's descriptions.
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.
- Figure-embedded text (axis labels, legends) is interleaved into the prose by PDF extraction (e.g., Figure 1's PCA labels), so figure-level readings rely on captions; all numbers cited in this review come from body text, captions, and Table 1.
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.