The critical role of evaluation metrics in handling missing data in machine learning

The presence of missing data in machine learning (ML) datasets remains a major challenge in building reliable models. This study explores various strategies to handle missing data and provides a framework to evaluate their effectiveness. The research focuses on commonly used techniques such as zero-filling, deletion, and imputation methods, including mean, median, mode, regression, k-nearest neighbors (KNN), and flagging. To assess these methods, a detailed evaluation framework is proposed, considering factors such as data completeness, model performance, stability, bias, variance, robustness to new data, computational efficiency, and domain-specific needs. This comprehensive approach allows for a thorough comparison of methods, helping to identify the most suitable technique for specific datasets and tasks. The findings highlight the importance of considering the unique features of the dataset and the goals of the analysis when choosing a method. While basic techniques like deletion and zero-filling may be effective in some cases, advanced imputation methods often preserve data quality and improve model accuracy. By applying the proposed evaluation criteria, researchers and practitioners can make better decisions on handling missing data, leading to more accurate, reliable, and adaptable ML models. (c) 2025 The Authors. Published by IASE. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).