A Comparative Evaluation of Imputation Methods for Support Vector Regression under Missing at Random

Abstract

Handling missing data remains a fundamental challenge in the context of Support Vector Regression (SVR). This study aims to compare the performance of five missing data imputation methods (Mean, Median, KNN, MICE, MissForest) and their impact on SVR accuracy, These methods range from single-value imputation (Mean, Median) to similarity-based (KNN) and relationship-based (MICE, MissForest) approaches. The study employed these methods through two analytical tracks using a Monte Carlo simulation study, which varied the noise level, number of predictors, and sample size, alongside the real-world Boston Housing Dataset with missingness rates of 5%, 10%, 20%, 30% applied in both tracks, and performance was evaluated using RMSE, MAE, and R². Results showed that increasing missingness negatively affects SVR accuracy in both analytical tracks. MICE and MissForest consistently outperformed the traditional methods: mean and median imputation were the weakest across every simulation condition and the real-data application, while KNN performed competitively, even outperforming MICE on the real data at higher missingness rates. MissForest excelled at low-to-moderate noise levels and MICE at high noise. Results showed also that increasing sample size improves the performance for all methods.