การตัดเล็มอย่างอ่อนสำหรับต้นไม้ตัดสินใจโดยการใช้แบ็กพรอพาเกชันนิวรอลเน็ตเวิร์ก / ก้องศักดิ์ จงเกษมวงศ์ = Soft pruning for decision tree using the backpropagation neural network / Kongsak Chongkasemwongse
Decision tree learning is a machine teachnique that is in widespread use nowadays. When we construct a decision tree from some data, especially noisy data, the obtained tree may overfit the data and may be very large. This result makes the tree not perform well on new data. The commonly used technique for preventing the overfitting is to prune the decision trees. Many methods for decision tree pruning have been proposed, and all of them remove some nodes from the tree to reduce its size. However, some removed nodes may have a significance level or some contribution in classifying new data. Instead of absolutely removing nodes, this thesis presents a new method that gives weights to nodes according to their significance. We call this method "soft pruning". The significance level or the weight of a node is determined by a backpropagation neural network. We run experiments on twenty domains to compare our method with error-based pruning that is one of the most effective method for free pruning. The results demonstrate that our method outperforms error-based pruning.