Feature Selection Based on a Shallow Convolutional Neural Network and Saliency Maps on Metagenomic Data

Abstract

In recent years, personalized medicine has been discovered by scientists to improve existing curative methods. These studies are mainly performed on metagenomic datasets which is the large dataset related to many human diseases, especially genetic data. The development of machine learning models and related algorithms has enabled us to speed up computation and improve disease diagnosis accuracy. However, due to the large dataset and the rather complicated processing of the data, we encountered certain difficulties. Therefore, we propose an approach to the task of selecting features based on the explanatory model. This approach is made up of proposing a small set of features from the original, implemented with Explanations with Saliency Maps. The results exhibit better performances comparing to random feature selection. Explanations generated by Saliency Maps have provided a promising method in selecting features and are expected to apply in practical cases.