Missing Data Imputation Using Data Generated By GAN

Abstract

Missing data is a common and challenging problem that arises in many research domains and led to the complication of data analysis. Therefore, handling missing data is a necessity as proposed in many previous studies. In this paper, we proposed two methods to impute missing numerical datasets based on generated data by GAN and determine the imputed values using Euclidian distance. In various missing percentages, we evaluated the imputation accuracy of all methods using MAE and RMSE tests. The proposed methods randomGAN and meshGAN produce the best imputation accuracy in 2 out of 4 datasets against three compared methods: SimpleImputer, IterativeImputer, and KNNimputer.