Publication: Multi-objective evolutionary feature selection for ensemble learning with random forests in time series forecasting
Authors
Espinosa, Raquel ; Sánchez Carpena, Gracia ; Palma Méndez, José Tomás ; Jiménez Barrionuevo, Fernando
item.page.secondaryauthor
Facultades de la UMU::Facultad de Informática
item.page.director
Publisher
Elsevier
publication.page.editor
publication.page.department
DOI
https://doi.org/10.1016/j.swevo.2025.102211
item.page.type
info:eu-repo/semantics/article
Description
Abstract
Time series forecasting is fundamental in numerous domains, including finance, healthcare, energy, and environmental monitoring. However, the high dimensionality of feature spaces can lead to overfitting and reduced interpretability, making feature selection a crucial preprocessing step. This paper proposes a multiobjective evolutionary algorithm for feature selection in time series forecasting, designed to enhance predictive accuracy while improving generalization. The method partitions the dataset, associating each partition with an objective function in the optimization process. By independently selecting relevant feature subsets, it generates a Pareto front of Random Forest models, each trained on a distinct subset of features. These models are then aggregated into a stacking-based ensemble framework, effectively balancing feature relevance and diversity. Additionally, we introduce a feature importance measure based on selection frequency in the non-dominated solutions of the optimization process. To validate our approach, we conduct experiments on real-world forecasting tasks, including air quality prediction in southeastern Spain and Italy and oil temperature forecasting in industrial applications. We also evaluate performance on synthetic datasets of increasing complexity, systematically varying instances, features, seasonality, noise, and trends. The proposed method is compared against conventional Random Forest, a wrapper-based feature selection method with a multiobjective evolutionary search strategy, and several state-of-the-art embedded feature selection techniques for time series forecasting. The results demonstrate that our approach significantly improves forecasting accuracy while mitigating overfitting. By integrating multi-objetive evolutionary optimization, random forest, ensemble learning, and a novel feature importance measure, our method offers a robust, interpretable, and effective feature selection for time series forecasting applications.
publication.page.subject
Citation
Swarm and Evolutionary Computation, 2025, Vol. 99 : 102211
item.page.embargo
Collections
Ir a Estadísticas
Este ítem está sujeto a una licencia Creative Commons. http://creativecommons.org/licenses/by-nc-nd/4.0/