Assessing the Stability of Classification Trees at Different Cost Ratios in Cost-Sensitive Learning

• Activity Number: 652
• Type: Contributed
• Date/Time: Thursday, August 2, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Hongjuan Liu*+ and Lisa Dooley
• Companies: and Johnson & Johnson
• Keywords: Classification Trees ; Stability ; Imbalanced Datasets ; Cost Sensitive Learning ; Altered Prior ; Delta Method

Abstract

Classification tree is well known for its instability in that a small perturbation in the training dataset can result in a significant different tree structure. Research on the stability of the classification tree has generally been concerned with the instability of class prediction. Little work has been done on examining their robustness with respect to the tree structure. We analyze the causes of the instability of the tree and study an issue of the tree stability in the realm of imbalanced data sets - the relation between the tree stability (i.e., split stability) and the misclassification cost and the class prior. We also discuss experimentally the performance of the predictive accuracy at different cost ratios of the classification tree. The result reveals that certain misclassification cost ratio produces trees that were more structurally stable and predictive accurate than other ratios. This suggests that we need to take into account the influence of the misclassification cost in the stability and predictive performance when we use the cost sensitive learning to target the practically and theoretically important imbalance classification problem.