Gaussian processes for missing value imputation

Bahram Jafrasteh, Daniel Hernández-Lobato, Simón Pedro Lubián-López, Isabel Benavente-Fernández

Abstract: Highlights•A novel approach based on chained GPs, named MGP, is introduced for imputing missing values.•The method outputs a predictive distribution for each missing value in the dataset.•The final model can be trained simultaneously and can be scaled to large data sets.•Better performance of MGP than state-of-the-art methods for missing value imputation.