วิธีการตัดเล็มอย่างอ่อนสำหรับต้นไม้ตัดสินใจโดยใช้ฟัซซิฟิเคชัน / วิฆเนศ ทองมี = An approach of soft pruning for decision trees using fuzzification / Wikanes Thongmee
Decision trees have been widely and successfully used in machine learning nowadays. However, they have suffered from the overfitting problem in noisy domain which causes too many details in decision trees and reduces their accuracies. This problem has been remedied by decision tree pruning. Many methods for decision tree pruning have been proposed with the same basis of removing some nodes from the tree completely. However, the removed nodes may have some significant roles in data classification. Thus, a technique of decision tree pruning without node removal, called soft-pruning, has been proposed. Soft-pruning gives weights to nodes according to their significances which are determined by a backpropagation neural network. This thesis proposes a novel fuzzy method for soft-pruning decision trees called fuzzy soft-pruning. This method uses fuzzy membership functions to represent the decision of each node in the soft-pruning process instead of sharp boundary decision. The sigmoid membership functions were used for continuous attributes, while triangular membership functions were used for discrete attributes to give some levels of uncertainty to values around the threshold of decision. Experimental results on seventeen multi-class domains demonstrate that the novel method outperforms both C4.5's error-based pruning and soft-pruning.