ProToMEx: Rapid, Interpretable Explanations via Structured Representations
Athina Georgara, Adarsh Valoor, Sarvapali D. Ramchurn
cs.LG · cs.AI
Quick verdict
Uses LDA topics to explain black-box classifiers fast; idea is promising but validation is circular (synthetic only) and fidelity metrics one-sided.
Summary
ProToMEx is a model-agnostic post-hoc explainer that repurposes Latent Dirichlet Allocation for tabular classifiers. Offline, it samples synthetic instances from the feature domain, labels them with the black-box model (a 5-layer ReLU network), discretises each feature into class-conditional intervals, and renders every instance-label pair as a weighted bag-of-words 'document' whose feature-value tokens are repeated according to how characteristic they are of the predicted class. The learned topics serve as global prototypes per class and, via cheap single-document inference, as local explanations that can expose several co-existing reasons. Across 9 synthetic and 4 Kaggle datasets with 100 random seeds, ProToMEx recovers ground-truth generative patterns with average topic similarity 0.62-0.82 (Table 4), misses fewer than 3 SHAP/LIME-influential features for most instances (Tables 6-7), and explains an instance in ~1 ms versus ~23-37 ms for SHAP, breaking even after 90-2600 instances. Limitations are substantial: ground truth exists only on synthetic data whose structure mirrors LDA's inductive bias; fidelity is scored one-sidedly as coverage of baseline features while ProToMEx lists roughly twice as many; key hyperparameters (topic count K, δ, |B|, ε) are unreported; the headline '30-40x faster' exceeds the ~19-30x SHAP ratios implied by Table 5; and the 100-seed averages carry no variance or significance tests. Code is promised upon acceptance.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 4
● 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 translation of instance-label pairs into weighted bag-of-words documents lets LDA topics serve as both global per-class explanations and cheap local explanations that can disentangle multiple co-existing reasons for one prediction (§4).
- On synthetic data, detected topics match ground-truth generative patterns with average similarity 0.62–0.82 across the nine settings (Table 4).
- ProToMEx misses fewer than 3 SHAP/LIME-influential features in the majority of test instances across all four public datasets (Tables 6–7).
- Local explanations are dramatically cheaper after amortisation: ~1 ms per instance vs ~23–37 ms for SHAP, with offline cost recouped after 90–2600 explained instances depending on dataset (Tables 3, 5; §5.3).
- To the authors' knowledge this is the first successful adaptation of probabilistic topic models to post-hoc explanation of arbitrary black-box classifiers (§1).
Strengths
Genuinely novel amortisation angle
One offline LDA fit yields both global per-class topics and ~1 ms local inference (§4.2); §5.3 honestly reports break-even points (2600/1225/125/90 instances on Loan/Wine/Air/Lungs) rather than quoting only per-instance speedups.
Fully specified algorithms
Algorithms 1–2 give a complete, implementable pipeline: class-conditional PMFs μ over δ intervals, rank-based |B|-bin token weighting, label token repeated |B| times, and topic thresholds ε/ε' for local/global explanations.
Addresses a real gap in attribution lists
Broad, repeated experimental protocol
Candour about limits
Weaknesses
Circular ground-truth validation
One-sided fidelity metric with over-generation
Unreported hyperparameters
Speedup claims do not reconcile with the tables
Internal inconsistencies in dataset descriptions
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
ProToMEx (arXiv 2609.04265, cs.LG/cs.AI; Georgara, Valoor, Ramchurn, Southampton) proposes Probabilistic Topic Models — concretely LDA — as post-hoc explainers for black-box classifiers. Four claims: (1) a tabular-to-corpus translation lets LDA topics act as global per-class prototypes and cheap local explanations, uniquely able to expose multiple co-existing reasons (§1, §4); (2) fidelity comparable to SHAP/LIME (abstract; Tables 6–7); (3) amortised local explanation ~30–40× faster than SHAP/LIME (abstract; Tables 3, 5); (4) first adaptation of PTMs to post-hoc explanation of any black-box classifier (§1, hedged). Claims 1 and 4 are plausible; claims 2 and 3 are where the evidence is weakest.
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 ProToMEx in one sentence?
ProToMEx is a post-hoc explainability method that trains an LDA topic model on documents built from a classifier's (feature-value, label) pairs, so that the learned topics act as reusable 'reasons' — global prototypes per class and fast local explanations for individual predictions.
How much faster is ProToMEx than SHAP?
In the paper's experiments, explaining one instance takes about 1 ms with ProToMEx versus roughly 23–37 ms with SHAP on four public tabular datasets (about 19–30×), after a one-off corpus-building and LDA-training cost that is recouped after explaining 90–2600 instances; the abstract's '30–40× faster than SHAP and LIME' headline is more generous than its own tables.
Does ProToMEx replace SHAP or LIME?
No — it trades granularity for throughput: it outputs topic-level feature-set explanations, its reported fidelity is measured only as coverage of the features SHAP/LIME flag as influential, and it lists roughly twice as many features as those baselines, so precision of its explanations is not established.
Is ProToMEx's code or data available?
The paper says the code will be publicly available 'upon acceptance' (so not yet released), while the four public datasets are Kaggle datasets linked through a bit.ly shortcut in the paper.
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.
- Table 3 and Table 5 headers are garbled in the extraction (column labels interleaved, e.g., 'Corpus ProToMEx Explain Instance (×10−2 sec) Gen (sec) training (sec)'); the SHAP-vs-LIME column assignment in Table 5 was reconstructed from Fig. 2 values and break-even arithmetic, so those two columns could conceivably be swapped.
- Mathematical notation is partially garbled in extraction (the η scaling factor in Eq. 1 and the expectation symbol in the μ definition); the similarity metric was read by best effort.
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.