Constructing optimal forests of decision trees

Details

Supervisors

Verhaeghe, Hélène ; Nijssen, Siegfried

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques

Abstract

In this master thesis we aim to show a new method of building forests of decision trees using methods borrowed from optimisations. We first show that a naive method based on smarter decision trees doesn't work. The reasons for this are then explored in detail and some alternative improvements are given and analysed. We then focus on showing how a more in depth algorithm based on column generation performs in comparison to state of the art methods that work similarly, and show cases where it works better than those. We also give further improvements that can be added in the future, and see which bottlenecks still need to be solved for this method to be applicable in practice.