Evaluation of the pruning impact on Fuzzy C4.5
Marcos E. Cintra, Maria Carolina Monard, Heloisa A. Camargo, Caixa Postal · 2010
Decision trees are a consolidated method in machine learning for classification and regression tasks, and several algorithms have been proposed for their induction. C4.5 is a well-known decision tree algorithm that uses the information gain and entropy measures when deciding on the importance of the features. This way, decision trees also perform an attribute selection during the induction of the tree. Fuzzy decision trees have also been proposed in the literature and have advantages related to the discretization of continuous attributes into fuzzy sets, besides allowing the use of different reasoning methods, since the fuzzy decision trees compute a degree of compatibility for each triggered rule. To avoid overfitting, a pruning process may be performed. In this paper we analyse the impact of different pruning rates when generating fuzzy trees. Results, using 10 datasets, show that although the difference in the classification is small, the interpretability of the generated models vary significantly with the variation of pruning, allowing for a compromise between accuracy and interpretability.