A Discrete Missing Data Imputation Method Based on Improved Multi-layer Perceptron

Data have become an important factor of production in various fields. While there are many missing values due to some irresistible reasons, which may lead to incorrect results and conclusions. A set of imputation methods have been proposed for filling continuous data, while only a few works focus on discrete missing data. In this paper, we propose an improved multi-layer perceptron (IMLP) imputation method based on the momentum gradient descent (MGD) algorithm to impute the discrete missing data via the one-hot encoding technique. To verify the performance of this method, comparisons have been made of mode, K-nearest neighbor, and auto-encoder imputation methods on different missing patterns and missing rates. The simulation result shows that the IMLP has a good imputation performance on discrete data with a high missing rate.