Fecha de lectura

09/10/2026

Director

Juan Antonio Botía Blaya; José Tomás Palma Méndez

Grupo de investigación

Sistemas Inteligentes y Telemática

Departamento

Ingeniería de la Información y las Comunicaciones

Centro

Facultad de Informática

Organismo

Universidad de Murcia

Ciudad

Murcia

País

España

Abstract

This thesis addresses three challenges that co-occur in real-world machine learning: severe class imbalance, concept drift, and the reliability of model explanations. It proposes FILM (Framework for Imbalanced Learning Machines), validated on a real-world COVID-19 clinical registry from the Murcia Health Service and on public benchmark datasets. For static problems, it introduces BAIC, a metric that integrates performance indicators weighted by their bias towards the minority class proportion, and IPIP, an ensemble algorithm built on balanced partitions without synthetic data. Both are extended to data streams through Adaptive BAIC and IPIP-CD, a passive adaptation strategy that achieves the best performance in six of eight evaluated streams. The thesis also shows that class imbalance alters the magnitude, variance and stability of SHAP explanations, and proposes the 3P approach, which combines data, model and explanation perspectives to detect and interpret concept drift.

Visualizar Tesis