Bridging Data Gaps: A Comparative Study of Different Imputation Methods for Numeric Datasets

In machine learning and data science, filling in the missing values in data is crucial to maintaining the accuracy and integrity of predictions and analyses. Despite the variety of imputation methods available, there is a lack of comprehensive benchmarks comparing classical and modern approaches, leaving a significant gap in understanding the most effective strategies for handling missing values. This work attempts to close this gap by conducting several experiments on various datasets with different levels of missing values. We compare approaches like SimpleImputer (Mean), Iterative Imputer, kNN Imputer, MissForest Imputer, and Gaussian Copula Imputer. We illustrate the superior performance of the Gaussian Copula-based imputation approach across varied percentages of missing values using the Kolmogorov-Smirnov metric and Adversarial Validation Compliment score. Our findings contribute to a deeper understanding of imputation methods and provide practical insights for researchers and practitioners, aiding in informed decision-making during data preprocessing and analysis.