Partially supervised feature selection using tree ensemble approaches

Details

Supervisors

Dupont, Pierre

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques

Abstract

This work offers a mechanism to perform feature selection using the benefits of having a priori partial knowledge about the relevance of some variables in a given context. It extends the embedded feature selection method of random forests by increasing the probability of the relevant features to be evaluated by Gini index at each node. We show that using the a priori knowledge in this way will deviate the embedded mechanism of variables selection towards the favored ones in a way that the overall selected variables will be indeed relevant. This kind of favoring has some drawbacks like overusing the favored variables, but it can be avoided by increasing the number of variables from which the selection is performed.