Phenotype prediction with semi-supervised classification trees

Abstract

In this work, we address the task of phenotypic traits prediction using methods for semi-supervised learning. More specifically, we propose to use supervised and semi-supervised classification trees as well as supervised and semi-supervised random forests of classification trees. We consider 114 datasets for different phenotypic traits referring to 997 microbial species. These datasets present a challenge for the existing machine learning methods: they are not labelled/annotated entirely and their distribution is typically imbalanced. We investigate whether approaching the task of phenotype prediction as a semi-supervised learning task can yield improved predictive performance. The results suggest that the semi-supervised methodology considered here is especially helpful when using single trees, especially when the amount of labeled data ranges from 20 to 40%. Similar improvements can be seen when the presence of the phenotype is very imbalanced.