On Missing Values and the Imputation in Learning Neural Networks

Effective handling of missing values and imputation is critical in data-driven learning. Various approaches are available, including imputation, deletion, and direct training with incomplete data. Each method has its own set of advantages and disadvantages, impacting learning performance differently. This paper aims to evaluate the performance of a neural network model using two complete data sets, where different percentages of missing values are later introduced and imputed with zero and mean methods. We compare model performance with complete, imputed data and present two adapted versions to handle missing values. The experiments involve introducing missing values into the data sets and applying zero and mean imputation methods. Additionally, we introduce two adapted models, namely Masked and MissNN, specifically designed to handle missing values. Our experiments demonstrate that the adapted models, Masked and MissNN, exhibit greater consistency and minor spread in performance, achieving results comparable to training with complete data.