First-Mover Bias in Gradient Boosting Explanations: Mechanism, Detection, and Resolution

Fuente: Zenodo
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Caraker, Drake, Arnold, Bryan, Rhoads, David
Format: Recurso digital
Langue:anglais
Publié: Zenodo 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866901957301501952
author Caraker, Drake
Arnold, Bryan
Rhoads, David
author_facet Caraker, Drake
Arnold, Bryan
Rhoads, David
contents <p>Abstract<br><br>We isolate and empirically characterize first-mover bias—a path-dependent concentration<br>of feature importance caused by sequential residual fitting in gradient boosting—as a specific<br>mechanistic cause of the well-known instability of SHAP-based feature rankings under mul-<br>ticollinearity. When correlated features compete for early splits, gradient boosting creates a<br>self-reinforcing advantage for whichever feature is selected first: subsequent trees inherit modified<br>residuals that favor the incumbent, concentrating SHAP importance on an arbitrary feature<br>rather than distributing it across the correlated group. Scaling up a single model amplifies this<br>effect—a Large Single Model with the same total tree count as our method produces the worst<br>explanations of any approach tested.<br><br>We demonstrate that model independence is sufficient to resolve first-mover bias in the linear<br>regime, and remains the most effective mitigation under nonlinear data-generating processes.<br>Both our proposed method, DASH (Diversified Aggregation of SHAP), and simple seed-averaging<br>(Stochastic Retrain) restore stability by breaking the sequential dependency chain, confirming<br>that the operative mechanism is independence between explained models, not any particular<br>aggregation strategy. At ρ = 0.9, both methods achieve stability = 0.977, while the standard<br>single-best workflow degrades to 0.958 and the Large Single Model to 0.938. On the Breast<br>Cancer dataset, DASH improves stability from 0.53 to 0.93 (+0.40) over the standard Single<br>Best, and from 0.32 to 0.93 (+0.61) over the training-budget-matched Single Best (M =200).<br><br>DASH additionally provides two novel diagnostic tools—the Feature Stability Index (FSI)<br>and Importance-Stability (IS) Plot—that detect first-mover bias without ground truth, enabling<br>practitioners to audit explanation reliability before acting on feature rankings. Software and<br>reproducible benchmarks are available at https://github.com/DrakeCaraker/dash-shap.<br><br>Keywords: first-mover bias, SHAP, feature importance, multicollinearity, model independence,<br>gradient boosting, explainability, Rashomon effect</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19060133
institution Zenodo
language eng
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle First-Mover Bias in Gradient Boosting Explanations: Mechanism, Detection, and Resolution
Caraker, Drake
Arnold, Bryan
Rhoads, David
feature importance
multicollinearity
explainability
interpretable machine learning
XAI
ensemble explainability
diversity-aware aggregation
explanation aggregation
feature stability
XGBoost
SHAP
DASH
machine learning
<p>Abstract<br><br>We isolate and empirically characterize first-mover bias—a path-dependent concentration<br>of feature importance caused by sequential residual fitting in gradient boosting—as a specific<br>mechanistic cause of the well-known instability of SHAP-based feature rankings under mul-<br>ticollinearity. When correlated features compete for early splits, gradient boosting creates a<br>self-reinforcing advantage for whichever feature is selected first: subsequent trees inherit modified<br>residuals that favor the incumbent, concentrating SHAP importance on an arbitrary feature<br>rather than distributing it across the correlated group. Scaling up a single model amplifies this<br>effect—a Large Single Model with the same total tree count as our method produces the worst<br>explanations of any approach tested.<br><br>We demonstrate that model independence is sufficient to resolve first-mover bias in the linear<br>regime, and remains the most effective mitigation under nonlinear data-generating processes.<br>Both our proposed method, DASH (Diversified Aggregation of SHAP), and simple seed-averaging<br>(Stochastic Retrain) restore stability by breaking the sequential dependency chain, confirming<br>that the operative mechanism is independence between explained models, not any particular<br>aggregation strategy. At ρ = 0.9, both methods achieve stability = 0.977, while the standard<br>single-best workflow degrades to 0.958 and the Large Single Model to 0.938. On the Breast<br>Cancer dataset, DASH improves stability from 0.53 to 0.93 (+0.40) over the standard Single<br>Best, and from 0.32 to 0.93 (+0.61) over the training-budget-matched Single Best (M =200).<br><br>DASH additionally provides two novel diagnostic tools—the Feature Stability Index (FSI)<br>and Importance-Stability (IS) Plot—that detect first-mover bias without ground truth, enabling<br>practitioners to audit explanation reliability before acting on feature rankings. Software and<br>reproducible benchmarks are available at https://github.com/DrakeCaraker/dash-shap.<br><br>Keywords: first-mover bias, SHAP, feature importance, multicollinearity, model independence,<br>gradient boosting, explainability, Rashomon effect</p>
title First-Mover Bias in Gradient Boosting Explanations: Mechanism, Detection, and Resolution
topic feature importance
multicollinearity
explainability
interpretable machine learning
XAI
ensemble explainability
diversity-aware aggregation
explanation aggregation
feature stability
XGBoost
SHAP
DASH
machine learning
url https://doi.org/10.5281/zenodo.19060133