Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Hyperparameter Sensitivity of SMOTE and its Effects on Model Explainability

Holmin, Hugo LU (2026) STAN40 20261
Department of Statistics
Abstract
This study investigates how varying hyperparameters of the Synthetic Minority
Over-sampling Technique (SMOTE) affects the stability of model explanations produced by SHapley Additive exPlanations (SHAP) and Local Interpretable Modelagnostic Explanations (LIME). By using an eXtreme Gradient Boosting (XGBoost)
classifier trained on 11 medical datasets with varying imbalance ratios, four SMOTE
parameters are evaluated: sampling strategy, number of nearest neighbors k, interpolationdistribution, and the introduction of Gaussian noise. Explainability stability is measured as the percentage change in mean absolute SHAP and LIME values compared to a baseline model trained without oversampling.

Results show that sampling strategy has the... (More)
This study investigates how varying hyperparameters of the Synthetic Minority
Over-sampling Technique (SMOTE) affects the stability of model explanations produced by SHapley Additive exPlanations (SHAP) and Local Interpretable Modelagnostic Explanations (LIME). By using an eXtreme Gradient Boosting (XGBoost)
classifier trained on 11 medical datasets with varying imbalance ratios, four SMOTE
parameters are evaluated: sampling strategy, number of nearest neighbors k, interpolationdistribution, and the introduction of Gaussian noise. Explainability stability is measured as the percentage change in mean absolute SHAP and LIME values compared to a baseline model trained without oversampling.

Results show that sampling strategy has the most notable impact on explainability,
with higher oversampling ratio producing greater differences in SHAP values
from the baseline. Dataset imbalance ratio is identified as a consistent factor, with highly imbalanced datasets exhibiting larger percentage shifts in both SHAP and LIME across all parameter configurations. In contrast, changes to the interpolation distribution, k-neighbors, and the addition of Gaussian noise do not produce systematic differences in either explainability or predictive metrics. These findings confirm and extend prior work suggesting that SMOTE distorts feature attributions, and indicate that this distortion is mostly sensitive to the degree of oversampling applied. Practitioners are advised to treat sampling strategy with particular care when model explainability is a priority, and to be cautious about drawing conclusions from SHAP or LIME values derived from models trained on data augmented by SMOTE. (Less)
Please use this url to cite or link to this publication:
author
Holmin, Hugo LU
supervisor
organization
course
STAN40 20261
year
type
H1 - Master's Degree (One Year)
subject
keywords
SMOTE, XAI, SHAP, LIME
language
English
id
9231494
date added to LUP
2026-06-08 14:38:15
date last changed
2026-06-08 14:38:15
@misc{9231494,
  abstract     = {{This study investigates how varying hyperparameters of the Synthetic Minority
Over-sampling Technique (SMOTE) affects the stability of model explanations produced by SHapley Additive exPlanations (SHAP) and Local Interpretable Modelagnostic Explanations (LIME). By using an eXtreme Gradient Boosting (XGBoost)
classifier trained on 11 medical datasets with varying imbalance ratios, four SMOTE
parameters are evaluated: sampling strategy, number of nearest neighbors k, interpolationdistribution, and the introduction of Gaussian noise. Explainability stability is measured as the percentage change in mean absolute SHAP and LIME values compared to a baseline model trained without oversampling.

Results show that sampling strategy has the most notable impact on explainability,
with higher oversampling ratio producing greater differences in SHAP values
from the baseline. Dataset imbalance ratio is identified as a consistent factor, with highly imbalanced datasets exhibiting larger percentage shifts in both SHAP and LIME across all parameter configurations. In contrast, changes to the interpolation distribution, k-neighbors, and the addition of Gaussian noise do not produce systematic differences in either explainability or predictive metrics. These findings confirm and extend prior work suggesting that SMOTE distorts feature attributions, and indicate that this distortion is mostly sensitive to the degree of oversampling applied. Practitioners are advised to treat sampling strategy with particular care when model explainability is a priority, and to be cautious about drawing conclusions from SHAP or LIME values derived from models trained on data augmented by SMOTE.}},
  author       = {{Holmin, Hugo}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Hyperparameter Sensitivity of SMOTE and its Effects on Model Explainability}},
  year         = {{2026}},
}