Handling Missing Values via a Neural Selective Input Model

Missing data represent an ubiquitous problem with numerous and diverse causes. Handling Missing Values (MVs) properly is a crucial issue, in particular in Machine Learning (ML) and pattern recognition. To date, the only option available for standard Neural Networks (NNs) to handle this problem has been to rely on pre-processing techniques such as imputation for estimating the missing data values, which limited considerably the scope of their application. To circumvent this limitation we propose a Neural Selective Input Model (NSIM) that accommodates different transparent and bound models, while providing support for NNs to handle MVs directly. By embedding the mechanisms to support MVs we can obtain better models that reflect the uncertainty caused by unknown values. Experiments on several UCI datasets with both different distributions and proportion of MVs show that the NSIM approach is very robust and yields good to excellent results. Furthermore, the NSIM performs better than the state-of-theart imputation techniques either with higher prevalence of MVs in a large number of features or with a significant proportion of MVs, while delivering competitive performance in the remaining cases. We demonstrate the usefulness and validity of the NSIM, making this a first-class method for dealing with this problem.