Evolutionary algorithms for learning ensembles of interpretable classifiers

• Ano de defesa: 2022
• Autor(a) principal: Cagnini, Henry Emanuel Leal
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Pontifícia Universidade Católica do Rio Grande do Sul; Escola Politécnica; Brasil; PUCRS; Programa de Pós-Graduação em Ciência da Computação
• Palavras-chave em Português: Evolutionary Algorithms; Ensemble Learning; Machine Learning; Interpretability; Classification; Supervised Learning; Regression; Algoritmos Evolutivos; Aprendizado de Máquina; Interpretabilidade; Classificação; Aprendizado Supervisionado; Regressão; CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: https://tede2.pucrs.br/tede2/handle/tede/10197

Classification is the machine learning task of categorizing instances into classes. There are several algorithms in the literature that perform classification, with varying degrees of success. For the most part, predictive performance was the pursued objective among practitioners and the academic community regarding the design of novel classification algorithms. More recently, however, interpretability has been gaining more and more attention. One area of machine learning that can benefit from increased interpretability is that of ensemble learning. Ensemble learning aims to reunite models that, when ensembled, can provide a high degree of predictive performance, even though the individual classifiers of the ensemble are often not much better at predicting classes than random guessing. Hence, the benefits are twofold: ensembles can improve predictive performance of interpretable (white-box) models that perform, on average, worse than black-box models; and the use of white-box models improves the interpretability of ensembles. In this thesis, we design two evolutionary algorithms (a powerful soft computing technique) to develop two ensemble learning methods, EDNEL and PUMA. PUMA learns ensembles of classifiers in a univariate strategy, assuming independence among variables, while EDNEL takes into account variable dependence through correlation analysis. However, in the thorough experimental analysis performed, we found that PUMA performs better than EDNEL with regards to average rank, whilst it is statistically equivalent to two well-established ensemble learning algorithms, Adaboost and Random Forests.