Scholarlink

<p>Missing data is a recurring difficulty in data–driven modelling, often reducing the reliability of predictive systems and weakening the relationships among key features. This study conducts a comprehensive comparison of twelve imputation techniques across two domains—Air Quality and Workforce analytics—under a controlled setting where 10% MCAR missingness is introduced into the six most influential numerical variables identified using SHAP. The evaluation integrates preprocessing, feature selection, missingness generation, imputation, statistical testing, and downstream classification to ensure a consistent and rigorous assessment framework.The experimental results show a clear advantage for ensemble gradient–boosting approaches. Across both large–scale datasets, the Ensemble XGB–LGBM and the proposed Ensemble All3 (XGBoost + LightGBM + CatBoost) achieved the strongest performance, with the lowest errors among all methods. In the Air-Large dataset, Ensemble All3 reached a MAE of 17.61 and RMSE of 44.90, while in the HR-Large dataset it obtained a MAE of 5.59 and RMSE of 16.42. These results were consistently superior to classical imputers such as Mean, Median, KNN, IterativeImputer, and MICE. Wilcoxon and Nemenyi statistical tests further confirmed the significance of these performance differences. SHAP-based error analysis also showed that ensemble models exhibit the lowest global error sensitivity, indicating more stable behaviour when handling complex feature interactions.Overall, the findings demonstrate that ensemble gradient–boosting imputers provide accurate, stable, and generalizable missing-value reconstruction across heterogeneous domains, making them a practical and interpretable choice for real-world data analysis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Comprehensive Evaluation of Missing Value Imputation Methods Using Machine Learning and Ensemble Models Across Multi-domain Datasets

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Farzana Tasnim,
  • Shefayatuj Johara Chowdhury,
  • Mohammad Saeed Hasan Chowdhury,
  • Ayesha Juleka,
  • Tanjim Mahmud,
  • Kaushik Deb,
  • Mohammad Shahadat Hossain

摘要

Missing data is a recurring difficulty in data–driven modelling, often reducing the reliability of predictive systems and weakening the relationships among key features. This study conducts a comprehensive comparison of twelve imputation techniques across two domains—Air Quality and Workforce analytics—under a controlled setting where 10% MCAR missingness is introduced into the six most influential numerical variables identified using SHAP. The evaluation integrates preprocessing, feature selection, missingness generation, imputation, statistical testing, and downstream classification to ensure a consistent and rigorous assessment framework.The experimental results show a clear advantage for ensemble gradient–boosting approaches. Across both large–scale datasets, the Ensemble XGB–LGBM and the proposed Ensemble All3 (XGBoost + LightGBM + CatBoost) achieved the strongest performance, with the lowest errors among all methods. In the Air-Large dataset, Ensemble All3 reached a MAE of 17.61 and RMSE of 44.90, while in the HR-Large dataset it obtained a MAE of 5.59 and RMSE of 16.42. These results were consistently superior to classical imputers such as Mean, Median, KNN, IterativeImputer, and MICE. Wilcoxon and Nemenyi statistical tests further confirmed the significance of these performance differences. SHAP-based error analysis also showed that ensemble models exhibit the lowest global error sensitivity, indicating more stable behaviour when handling complex feature interactions.Overall, the findings demonstrate that ensemble gradient–boosting imputers provide accurate, stable, and generalizable missing-value reconstruction across heterogeneous domains, making them a practical and interpretable choice for real-world data analysis.