Abstract: With the widespread application of deep learning,high-quality tabular data is crucial for model performance.However,missing values can severely disrupt the underlying data structure and distribution.Although numerous imputation methods exist,current research predominantly focuses on imputation accuracy,lacking a systematic evaluation of how imputation outcomes affect the interpretability of downstream models.This paper proposes a framework for evaluating missing value imputation methods from the perspective of model interpretability.Firstly,it explores the advantages of deep generative models in learning complex data distributions to generate high-quality imputed values.Next,it constructs various missing data scenarios and employs Shapley values as a core metric to quantitatively compare the impact of different imputation methods on model feature importance explanations.Experimental results demonstrate that:1)Deep generative models can effectively learn the sample distribution and excel at preserving data structure and informational integrity.2)There is no direct correlation between imputation accuracy and the stability of model explanations;the choice of imputation method significantly alters the final Shapley values.3)As the proportion of missing data increases,the differential impact of various imputation methods on model interpretability becomes more pronounced.This study reveals the latent impact of missing value imputation on model interpretability and provides empirical evidence and a new evaluation perspective for selecting appropriate imputation strategies in interpretability-critical scenarios.

Key words: Table data, Missing data, Deep generative model, Shapley value