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

Bryan Arnold; David Rhoads; Drake Caraker

arxiv: 2603.22346 · v2 · submitted 2026-03-22 · 💻 cs.LG · cs.AI

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

Drake Caraker , Bryan Arnold , David Rhoads This is my paper

Pith reviewed 2026-05-15 07:34 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords first-mover biasSHAP explanationsgradient boostingattribution stabilitymulticollinearitymodel independenceDASH

0 comments

The pith

Gradient boosting explanations suffer first-mover bias from sequential residual fitting, which independent models largely neutralize.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper identifies first-mover bias as a source of unstable SHAP feature attributions in gradient boosting, arising because each tree fits residuals left by prior trees and therefore privileges whichever features enter the sequence first when features are correlated. A single large model that matches the total tree count of an ensemble produces the worst reproducibility of any tested approach. Methods that break the sequential chain by training many independent models and aggregating their SHAP values, such as DASH or simple seed averaging, restore stability to approximately 0.977 even at rho=0.9. On the Breast Cancer dataset DASH raises stability from 0.376 to 0.925. The authors supply two diagnostic tools that detect the bias without ground truth and confirm the mechanism through formal ANOVA variance decomposition.

Core claim

First-mover bias is the path-dependent concentration of SHAP feature importance caused by the sequential residual-fitting process in gradient boosting; this bias produces poor attribution reproducibility under multicollinearity, but the effect is largely eliminated by training independent models whose SHAP values are aggregated.

What carries the argument

The sequential residual-fitting chain inside gradient boosting, which privileges early-selected features in the boosting sequence under multicollinearity and is broken by model independence in the DASH diversified aggregation procedure.

Load-bearing premise

The observed attribution instability is caused primarily by the sequential residual-fitting mechanism rather than by other factors such as tree depth, learning rate, or the specific SHAP approximation used.

What would settle it

Train gradient boosting models with simultaneous rather than sequential tree construction and measure whether SHAP reproducibility under high multicollinearity reaches the same levels achieved by DASH; matching levels would show the sequential mechanism is not the dominant driver.

read the original abstract

We identify first-mover bias -- path-dependent concentration of SHAP feature importance from sequential residual fitting in gradient boosting -- as a mechanistic contributor to attribution instability under multicollinearity. Scaling up a single model amplifies this effect: a Large Single Model matching our method's total tree count produces the poorest attribution reproducibility of any approach tested. We show that model independence largely neutralizes first-mover bias. Both DASH (Diversified Aggregation of SHAP) and simple seed-averaging (Stochastic Retrain) restore stability by breaking the sequential dependency chain. At rho=0.9, both achieve stability ~0.977, while Single Best degrades to 0.958 and LSM to 0.938. On Breast Cancer, DASH improves stability from 0.376 to 0.925 (+0.549), outperforming Stochastic Retrain by +0.063. Under nonlinear DGPs, the advantage emerges at rho>=0.7. DASH provides two diagnostic tools -- the Feature Stability Index and Importance-Stability Plot -- that detect first-mover bias without ground truth. A crossed ANOVA with formal F-statistics confirms the mechanism: DASH shifts variance from model-dominated (40.6%) to data-dominated (73.6%). Software at https://github.com/DrakeCaraker/dash-shap

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper flags a sequential first-mover effect in gradient boosting SHAP that hurts stability under multicollinearity and shows that independent-model averaging largely removes it.

read the letter

The main thing to know is that sequential residual fitting in gradient boosting creates path-dependent SHAP attributions when features correlate, and the authors show that switching to independent models (via DASH or even plain seed averaging) lifts stability from 0.938 to 0.977 at rho=0.9 while keeping total tree count fixed. On the Breast Cancer data the lift is even larger, from 0.376 to 0.925. They back this with a crossed ANOVA that moves variance from model-dominated to data-dominated and supply two simple diagnostics that work without ground truth. The GitHub link is a clear positive for anyone who wants to check the implementation. What the work does well is make the stability problem concrete and testable rather than hand-wavy, and the mitigation is low-cost enough that practitioners could adopt it quickly. The soft spot is that the experiments do not fully isolate the sequential mechanism. They match total trees across conditions but do not hold the boosting structure fixed while varying learning rate, max depth, or subsample ratio, so it remains possible that some of the reported gains come from generic ensemble variance reduction instead of specifically breaking path dependence. The abstract claims the first-mover bias is the primary driver, yet without those controls the causal story is suggestive rather than locked down. This paper is aimed at people who already use gradient boosting on tabular data and need stable explanations in the presence of correlated features. Anyone working on XAI for high-stakes applications will get immediate practical value from the diagnostics and the DASH recipe. It deserves a serious referee because the core observation is reproducible, the fix is cheap, and the remaining question about mechanism isolation is exactly the kind of thing reviewers can sharpen.

Referee Report

2 major / 2 minor

Summary. The manuscript identifies 'first-mover bias' -- path-dependent concentration of SHAP feature attributions arising from sequential residual fitting in gradient boosting -- as a contributor to explanation instability under multicollinearity. It shows that breaking sequential dependency via DASH (Diversified Aggregation of SHAP) or simple seed-averaging restores stability (e.g., ~0.977 at rho=0.9 vs. 0.938 for large single models), introduces diagnostics (Feature Stability Index, Importance-Stability Plot), and supports the mechanism with a crossed ANOVA showing variance shift from model-dominated (40.6%) to data-dominated (73.6%). Experiments match total tree count across methods and report gains on synthetic and Breast Cancer data.

Significance. If the results hold, the work supplies a mechanistic account and practical mitigation for a known source of instability in SHAP explanations of gradient boosting, a widely deployed class of models. The concrete stability numbers, formal ANOVA F-statistics, and public GitHub implementation constitute reproducible strengths that could improve trustworthiness of attributions in applied settings.

major comments (2)

[Experimental evaluation] Experimental protocol: total tree count is matched across Single Best, LSM, Stochastic Retrain, and DASH, yet no ablation holds the sequential residual-fitting structure fixed while varying learning rate, max_depth, or subsample ratio. Without this isolation, the claim that first-mover bias is the primary driver of the observed instability (rather than general hyperparameter effects) remains untested; the stability gains at rho=0.9 could therefore reflect ensemble variance reduction.
[ANOVA analysis] ANOVA variance decomposition: the reported shift (model variance 40.6% to data variance 73.6%) is presented as confirming the mechanism, but the manuscript does not detail how the crossed design controls for the specific SHAP approximation or data-generation process. This weakens the link between the statistical result and the sequential-fitting hypothesis.

minor comments (2)

[Abstract] The abstract states that the advantage emerges at rho >= 0.7 under nonlinear DGPs but does not specify the functional form of those DGPs or the exact multicollinearity construction.
Notation for stability metric and rho should be defined at first use in the main text for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and will revise the manuscript accordingly.

read point-by-point responses

Referee: [Experimental evaluation] Experimental protocol: total tree count is matched across Single Best, LSM, Stochastic Retrain, and DASH, yet no ablation holds the sequential residual-fitting structure fixed while varying learning rate, max_depth, or subsample ratio. Without this isolation, the claim that first-mover bias is the primary driver of the observed instability (rather than general hyperparameter effects) remains untested; the stability gains at rho=0.9 could therefore reflect ensemble variance reduction.

Authors: We agree that an ablation holding the sequential residual-fitting structure fixed while varying learning rate, max_depth, and subsample ratio would better isolate first-mover bias from general hyperparameter or ensemble effects. In the revised manuscript we will add this ablation, training sequential models across a grid of those hyperparameters and showing that attribution instability remains high under multicollinearity unless model independence is introduced. revision: yes
Referee: [ANOVA analysis] ANOVA variance decomposition: the reported shift (model variance 40.6% to data variance 73.6%) is presented as confirming the mechanism, but the manuscript does not detail how the crossed design controls for the specific SHAP approximation or data-generation process. This weakens the link between the statistical result and the sequential-fitting hypothesis.

Authors: We acknowledge that the manuscript provides insufficient detail on the crossed ANOVA controls. In the revision we will expand the methods section to describe the crossed factors explicitly, including how the design holds SHAP approximation parameters constant across model instances and crosses them with independent data-generation replicates, thereby linking the observed variance shift directly to the removal of sequential dependency. revision: yes

Circularity Check

0 steps flagged

No circularity: claims rest on direct empirical stability measurements and ANOVA

full rationale

The paper presents no derivation chain or first-principles prediction. All central claims (first-mover bias as contributor to instability, neutralization via model independence, diagnostic tools) are supported by explicit experimental comparisons of stability across retrainings, total-tree-matched baselines, and a crossed ANOVA variance decomposition. Stability metrics are computed from observed SHAP values on held-out retrainings rather than being defined in terms of any fitted parameter or self-referential equation inside the paper. No self-citation is load-bearing for the mechanism or results.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that gradient boosting's sequential residual fitting is the dominant source of attribution variance under multicollinearity, plus the empirical observation that model independence removes most of that variance. No free parameters are fitted to produce the stability numbers; the invented entity is the named bias itself.

axioms (1)

domain assumption Gradient boosting constructs an additive model by sequential residual fitting
Standard property of XGBoost, LightGBM, and similar implementations invoked throughout the abstract.

invented entities (1)

first-mover bias no independent evidence
purpose: Mechanistic explanation for path-dependent concentration of SHAP importance
Newly coined term used to label the sequential-fitting effect; no independent evidence outside the paper's experiments is supplied.

pith-pipeline@v0.9.0 · 5539 in / 1441 out tokens · 36112 ms · 2026-05-15T07:34:31.256814+00:00 · methodology

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

Core claim

What carries the argument

Load-bearing premise

What would settle it

discussion (0)