Análise de dinamicidade na seleção de parâmetros de comitê de classificadores

Abstract

Over the years, significant progress has been made in the realm of classifier ensembles research. Several methods to enhance their efficiency have been proposed, applicable to both homogeneous and heterogeneous ensemble structures, a key challenge lies in determining their structure (hyper-parameters). Basically, the ensemble structure selection can be done in two different ways, static and dynamic selection. Unlike static selection, which regardless of the parameters uses the same criteria to perform the classification, dynamic selection defines the ensemble structure for each test instance. Different dynamic selection methods have been proposed in the literature, mainly for ensemble members and dataset features, but little effort has been made to propose dynamic selection methods for combination methods, also known are fusion methods. Therefore, it is important to evaluate the impact of dynamic selection of combination methods or both (methods and members) in creating robust ensembles. This work proposes an exploratory analysis of the dynamic selection of the main parameters of an ensemble structure. To this end, three different scenarios are evaluated: Full static ensembles; Partially dynamic ensembles; and, Full dynamic ensembles. In order to analyze the dynamic scenarios, three dynamic fusion methods have been proposed. Each one focuses on a specific approach: one by region of competence, another by meta-learning, and the last by fuzzy hyperbox. Finally, an empirical analysis of these three scenarios was conducted on 30 datasets. The results of this research confirm that the dynamic selection of classifiers and combiners significantly improves the accuracy and adaptability of classifier ensembles. Fully dynamic methods demonstrated superior performance compared to partially dynamic and static approaches, standing out for their ability to select the best classifiers and fusion methods for each test instance. Thus, we can affirm that the results encourage the development of more efficient and scalable methods in the field of machine learning.