Aprendizado de máquina automatizado multiobjetivo para classificação multirrótulo

• Ano de defesa: 2025
• Autor(a) principal: Del Valle, Aline Marques
• Orientador(a): Cerri, Ricardo
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Português: Aprendizado de máquina automatizado; Classificação multirrótulo; Otimização multiobjetivo; Meta-aprendizado; Seleção de atributos
• Palavras-chave em Inglês: Automated machine learning; Multi-label classification; Multiobjective optimization; Meta-learning; Feature selection
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://hdl.handle.net/20.500.14289/23117

Multi-label classification is a challenging problem, and its resolution involves choosing a classification algorithm and its respective hyperparameters. However, finding the best algorithm-hyperparameter combination is not a trivial task. In this context, Automated Machine Learning (AutoML) emerges as a solution, automating the selection of the best algorithm and its hyperparameter configuration for a machine learning problem. For multi-label classification problems, selection occurs in a hyperspace designed for this problem category. This thesis addresses AutoML and multi-label classification. Despite advances in the field, mainly related to constructing multi-label classifiers, some issues remain open. In this work, we investigate three of these areas. We investigate whether employing multi-objective optimization in AutoML strategies for multi-label classification is feasible and efficient. To address this research question, we proposed the EMANUEL strategy (gEnetic Multi-objective strategy for the Automatic selectioN of mUlti-labEl cLassifiers), an AutoML strategy developed with the NSGA-II algorithm to find classifiers that maximize the Macro F-score while minimizing the model size. Our results showed that multi-objective optimization can find a diversified Pareto frontier, with classifiers that weight different objectives and perform competitively with other classifiers. Furthermore, we study techniques that reduce AutoML runtime without compromising results. To this end, we employ meta-learning using surrogate models to evaluate multi-label classifier algorithms and estimate the objective values of optimization, avoiding direct evaluations during AutoML strategy execution. Including these surrogate models in new versions of the EMANUEL strategy maintained the predictive quality of the final models while reducing the total AutoML runtime. Finally, we investigated the effect of including feature selection in the EMANUEL strategy. To this end, we extend the hyperspace of classification algorithms by adding multi-label feature selection algorithms. The Pareto frontiers resulting from the new version of EMANUEL contained classifiers that maintained performance relative to the objectives, typically using fewer features.