Autor: Ari Urkullu Villanueva

Tesis: Selection Procedures in Machine Learning: Scores, Estimations, and Reproducibility

Directores: Borja Calvo / Aritz Pérez

Día: 24 de marzo de 2025
Hora: 11:30h
Lugar: sala Ada Lovelace (Facultad de Informática)

Abstract:

"Supervised classification stands as a fundamental problem in machine learning, with the procedures for selecting features and models holding the highest importance in this context. In this domain, three key challenges emerge concerning model and feature selection: determining the suitable guiding quality measure for the selection process, accurately estimating this quality measure from the provided data, and analyzing the impact of uncertainty on the reproducibility of the selection process. This dissertation tackles three aspects related with the selection procedures used in supervised classification. In the field of classifier selection in data environments generated through crowdsourcing, three methods are proposed and studied to estimate the AUC when the underlying truth is not available. In the area of ranking-based feature selection, a statistical model is proposed to measure the reproducibility of ranking-based feature selection algorithms. Also in the area of ranking-based feature selection, alternative measures to traditional statistical tests are proposed, offering a better trade-off between reproducibility and performance in feature selection."