Prediction of missing values. Comparison of approaches

We compare four numerical methods for the prediction of missing values in different datasets [1]. These are 1) hierarchical maximum likelihood estimation (ℋ-MLE), and three machine learning (ML) methods, which include 2) k-nearest neighbors (kNN), 3) random forest, 4) Deep Neural Network (DNN). From ML best results (for considered datasets) were obtained by kNN method with (or seven) neighbors. On one dataset, MLE showed a smaller error than method, whereas, on another, was better. The requires lot linear algebra computations works fine almost all datasets. Its result can be improved taking threshold more accurate matrix arithmetics. To our surprise, well-known produces similar as ℋ-MLE worked much faster.