การจัดกลุ่มคุณลักษณะที่เหมาะสมเพื่อการสร้างต้นไม้ตัดสินใจที่มีประสิทธิภาพ

Abstract

The purpose of this study is to select a clustering method of features that is suitable for continuous data. The appropriateness of the clustering method, the number of clusters and the splitting method can increase the accuracy of decision tree. Two splitting methodologies used in this study are the two way split and the multi way split. Before the split process, 3 clustering algorithms are applied to each attribute. The clustering method in this study includes Expectation Maximization (EM), K-Means and Hierarchical. The numbers of clusters in the experimental are 2, 3, 4 and 5 clusters. Six standard dataset are used in the experimental including Iris dataset, Abalone dataset, Breast cancer Wisconsin dataset, Pima Indians diabetes dataset, Seeds dataset, and Ecoli dataset. The decision tree based on J48 algorithm creation is used for measurement the classification accuracy from the proposed splitting and clustering method in this study. The experimental results on Iris data set shows that decision tree using K-Means of 3 clusters, has the highest accuracy, that is 97.33%. The K-Means method tested with five other data sets finds that creating a decision tree by two ways split results in the higher accuracy over multi way split in 4 data set. Based on the comparing classification result between decision tree using the method in this study and the decision tree using ordinary J48, the proposed method results higher accuracy in the 5 data set out of 6 data set. In conclusion, clustering using K-Means with 3 clusters and creating a decision tree by two way split can improve the accuracy of decision trees that use J48 algorithm.