.png){kind=icon}
Return
Explainability of machine learning models under missing data
.png){kind=icon}
PRE
AI
DOI：10.1016/j.asoc.2026.115105.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• Impact of imputation on model explainability: Different imputation methods significantly influence Shapley values, affecting feature importance and interaction explanations derived from machine learning models. • Trade-off between prediction accuracy and explainability: A lower test prediction Mean Squared Error (MSE) does not necessarily correspond to a lower MSE in Shapley values, highlighting a trade-off between imputation accuracy and explainability preservation. • Dataset and missing rate dependency: The effects of imputation methods on Shapley values vary across datasets and become more pronounced as the missing data rate increases, emphasizing the need for context-specific imputation strategies. • Limitations of direct missing data handling in XGBoost: While XGBoost can handle missing data directly, it often distorts Shapley values more significantly compared to using imputed data, raising concerns about explainability. • Practical guidance for imputation method selection: This study offers insights and recommendations for selecting imputation methods based on dataset characteristics, missing rates, and the trade-off between accuracy and explainability.
Keywords:
Imputation methods
Shapley values
Model explainability
Missing data
Feature importance
Journal
IF:
6.6
Papers:
1.4W
Citations:
4.8W
Organization
Cited Papers
ShareSaveMissing data imputation with adversarially-trained graph convolutional networks.png)
NEURAL NETWORKS
OAAI
NEURAL NETWORKSIF6.3
OAAIShareSaveMissing Data Imputation for Industrial Time Series With Adaptive Median Iteration Based on Generative Adversarial NetworksIEEE SENSORS JOURNAL
IEEE SENSORS JOURNALIF4.5
PREAI
ShareSaveShareSaveShareSaveShareSaveShareSave
View more