FMMO: Detecting the Divergence Between Local Attribution and Global Drift
Muhammad Rehman Zafar, Ali El-Sharif, Naimul Khan
cs.LG
Quick verdict
FMMO pairs TreeSHAP, surrogate DT, and NannyML to catch drift hidden by stable local attributions; sensible integration, but no negative controls tested.
Summary
The paper proposes FMMO, a Framework for Model Monitoring and Observability that couples local XAI (TreeSHAP), a global surrogate decision tree, and NannyML-based monitoring (PCA-based domain-classifier drift detection plus confidence-based performance estimation, CBPE) to decide when deployed models need auditing while ground-truth labels are delayed. Across CES, CES-NA, COMPAS, German Credit, Communities & Crime, and a synthetic dataset, the authors inject Gaussian or uniform covariate noise into test sets and track top-k feature stability (Jaccard similarity and rank agreement) of RF and XGB explanations. The key finding is that TreeSHAP's aggregated attributions often stay identical (J=100%) while performance drops (e.g., Communities & Crime under uniform noise: ROC AUC 90.16%→81.09%), and on COMPAS the protected group's FPR jumps from 37.84% to 70.39% versus 10.17% to 36.17% for the privileged group — a fairness blind spot invisible to SHAP stability. The surrogate DT reacts more reliably, and FMMO unions all alerts. Main limitations: every configuration in Table 4 triggers an alert with no no-drift controls, so false-alarm behavior is unknown; thresholds are hand-set; no tuning, repetitions, confidence intervals, or significance tests; and the fairness evidence is retrospective since FPR requires labels.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● 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
- Aggregated TreeSHAP attributions can remain essentially unchanged (J=100%) under covariate drift that degrades performance, so local explanation stability is not a reliable drift or model-health signal.
- On COMPAS, injected drift raises the protected group's FPR from 37.84% to 70.39% while overall accuracy drops only ~4pp and SHAP stays stable — a fairness blind spot invisible to explanation monitoring.
- A decision-tree global surrogate reacts to covariate drift more reliably than TreeSHAP and complements multivariate drift detection (DC) and confidence-based performance estimation (CBPE).
- Unioning local XAI, surrogate, DC, and CBPE signals (FMMO) enables label-free flagging of deployed models for audit when ground truth is delayed or unavailable.
Strengths
Quantified fairness blind spot on COMPAS
Table 3 shows protected-group FPR rising 37.84%→70.39% (+32.55pp) vs privileged 10.17%→36.17% (+26.0pp) under high drift, with overall accuracy dropping only ~4pp, while TreeSHAP top-5 attributions are unchanged (Figures 4e/4f) — a concrete, memorable demonstration that explanation stability masks group-level harm.
Honest role separation across monitoring modules
The paper explicitly states what each signal cannot do (DC: presence-only, no feature insight; CBPE: no causal explanation; TreeSHAP: instance-level scope) and pairs two complementary stability metrics — Jaccard membership (J) and position-wise rank agreement (α) — rather than relying on one.
Includes a natural-drift condition, not only artificial noise
Candid self-assessment
Weaknesses
No negative controls; alert fires in all 22 configurations
No statistical rigor
Internal numeric inconsistencies
Circular drift calibration and unreported hyperparameters
Incremental novelty and suspicious bibliography
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
FMMO (arXiv:2609.06173v1, 5 Sep 2026) is a monitoring framework for tabular classifiers under covariate drift when ground-truth labels are delayed. It unifies four off-the-shelf signals: TreeSHAP local attributions, a decision-tree global surrogate (GSM), NannyML's PCA-based domain classifier (DC) for multivariate drift, and confidence-based performance estimation (CBPE). The paper makes three claims: (1) local attributions can remain stable under drift that degrades performance; (2) this stability hides a fairness blind spot, evidenced by protected-group FPR spikes on COMPAS; (3) unioning the signals (Algorithm 1: alert if J<β or α<ζ or DC or CBPE fires) flags models for audit. Claim (1) is well demonstrated; claim (2) is convincingly shown but only retrospectively; claim (3) is asserted rather than isolated — the abstract's statement that FMMO 'mitigates this fairness blind spot' overreaches, since the authors concede FPR is not a real-time signal (Section 'Drift in Sensitive Features') and FMMO's live outputs are generic drift alerts.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 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 FMMO?
FMMO (Framework for Model Monitoring and Observability) is a proposed monitoring pipeline that combines TreeSHAP local explanations, a decision-tree global surrogate model, and NannyML-based tools (a PCA-based domain classifier for multivariate drift detection and CBPE for confidence-based performance estimation). If any signal fires, it recommends auditing the deployed ML model — aimed at settings where ground-truth labels are delayed or missing.
Why can SHAP stability be misleading after data drift?
TreeSHAP explains individual predictions locally, and when attributions are averaged across a dataset they can remain nearly identical (Jaccard overlap of 100% in the paper's CES and COMPAS Gaussian-noise experiments) even while model accuracy drops and error rates for protected groups spike. Stable explanations therefore do not prove the model is still behaving as intended.
What datasets and models were used in the paper?
Five public/real tabular datasets — Consumer Expenditure Survey (CES and its 2017–2020 CES-NA year split), COMPAS recidivism, German Credit, Communities & Crime — plus a 25,000-row scikit-learn synthetic set. Random Forest and XGBoost classifiers were trained on original data and evaluated on noise-perturbed test data without retraining.
Does FMMO detect unfairness directly?
No. Its real-time signals are generic drift/performance alerts; the fairness evidence (protected-group FPR rising from 37.84% to 70.39% on COMPAS while SHAP stayed stable) was computed retrospectively with true labels, which the authors themselves acknowledge cannot be done in label-free deployment.
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 panels arrive as flattened axis-label text; visual judgments (e.g., threshold crossings in the DC/CBPE plots) rely on captions and in-text descriptions only.
- Table 4's rows were reconstructed from flattened extraction; individual cell values were cross-checked against the in-text narrative where available, but a few cell alignments (e.g., which k a J/α pair belongs to) depend on column ordering.
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.