A Hybrid GP-KNN Imputation for Symbolic Regression with Missing Values

In data science, missingness is a serious challenge when dealing with real-world data sets. Although many imputation approaches have been proposed to tackle missing values in machine learning, most studies focus on the classification task rather than the regression task. To the best of our knowledge, no study has been conducted to investigate the use of imputation methods when performing symbolic regression on incomplete real-world data sets. In this work, we propose a new imputation method called GP-KNN which is a hybrid method employing two concepts: Genetic Programming Imputation (GPI) and K-Nearest Neighbour (KNN). GP-KNN considers both the feature and instance relevance. The experimental results show that the proposed method has a better performance comparing to state-of-the-art imputation methods in most of the considered cases with respect to both imputation accuracy and symbolic regression performance.