Decision Tree with Better Ranking

AUC(Area Under the Curve) of ROC(Receiver Operating Characteristics) has been recently used as a measure for ranking performanceof learning algorithms. In this paper, wepresent a novel probability estimation algorithm that improves the AUCvalue of decision trees. Instead of estimating the probability at the single leaf wherethe example falls into, our methodaverages probability estimates fromall leaves of the tree. Thecontribution of each leaf is determined by the deviation in attribute values fromthe root to the leaf. Wedesign empirical experiments to verify that our newalgorithm outperforms C4.5 and its recent improvement C4.4 in AUC.Even though C4.4 with bagging outperforms our method in AUC,our method producesa single tree with interpretable results.

• Type: Conference
• Year: 2003
• Where: ICML
• Authors: Charles X. Ling, Robert J. Yan
• Keywords: ICML 2003, Machine Learning, Performanceof Learning Algorithms, Probability Estimation Algorithm, Single Leaf