Ranking the Rules and Instances of Decision Trees

Traditionally, decision trees rank instances by using the local probability estimations for each leaf node. The instances in the same leaf node will be estimated with equal probabilities. In this paper, we propose a hierarchical ranking strategy by combining decision trees and leaf weighted Näıve Bayes to improve the local probability estimation for a leaf node. We consider the importance of the rules, and then rank the instances fit in with the rules. Because the probability estimations based on Näıve Bayes might be poor, we investigate some different techniques which were proposed to modify Näıve Bayes as well. Experiments show that our proposed method has significantly better performance than that of other methods according to paired t-test. All results are evaluated by using AUC (Area under ROC Curve) instead of classification accuracy.