Preprocessing and ensemble approaches for singular problemsmonotonic and imbalanced classification

Resumen

This thesis focuses on two singular supervised problems: imbalanced classification and classification with monotonic constraints. This thesis aims to propose new solutions based on robust classifiers and preprocessing techniques for these two singular supervised problems. In essence, the main objective of this thesis is to design new solutions for these problems, both independently and together, and considering other singular data situations, such as the presence of class noise. As previously mentioned, these proposals follow two different approximations similar to the traditional approaches for singular problems: robust classifiers and preprocessing based techniques. These approaches aim at specific issues of imbalanced and monotonic classification: Design of robust classifiers for the singular problems of imbalanced and monotonic classification. This objective includes: -High levels of accuracy as well as monotonicity with a Random Forest classifier for monotonic classification. -Robust ensemble learning based on Switching according to the Nearest Enemy Distance (SwitchingNED) for highly imbalanced problems. -Great robustness to monotonic noise for monotonic classification with a Fuzzy kNN proposal aware to monotonic constraints and violations. Development of preprocessing-based techniques in imbalanced and monotonic classification. Within this paradigm, the following goals are enclosed: -To enhance the performance of SwitchingNED through its combination with different sampling techniques for imbalanced classification. -To mitigate the impact of skewed class distributions in problems with monotonic constraints thanks to new sampling techniques for monotonic imbalanced classification.