Quantile regression by optimal decision trees

Details

Supervisors

Siegfried Nijssen

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

Quantile regression is a field that has many interesting properties when compared to standard regression, including robustness with respect to datasets that present a lot of variability and have many outliers. In some contexts, these predictions need to be justifiable and hence, the models making said predictions need to be interpretable. This work will evaluate how optimal decision trees can be used to generate quantile predictions that are highly interpretable and how to analyze the trees obtained to not only explain the decisions for individual quantiles but to provide easily interpretable information on the whole conditional distribution of a dataset.