Indução de Árvore de Decisão utilizando Meta-Aprendizado
Caique Augusto Ferreira, José Augusto Baranauskas · 2023
Symbolic learning models stand out within the Machine Learning area because their representations are interpretable by humans. A feature of this model is very responsive to the set of examples used, which can result in a significant change in the model if there are small variations in the training set. The model combination strategy (ensembles) is presented as an alternative to improve the accuracy and stability of the models. The strategy consists of generating different models from the same training set and combining them into a single final model, usually through a voting process. An undesirable characteristic of the set strategy is the complexity of the final model, since it is formed by a set of models. In this research, an approach is proposed to induce a decision meta-tree based on the combination of decision trees from one (Random Forest). The experiments were performed on 150 datasets from different domains. The proposed approach applied to 43 sets of categorical data out of 150 analyzed, obtained a performance as good as a forest with 128 trees without statistically significant differences. This is an important result, considering the interpretability provided by a single decision tree as the resulting model.