249 – Estimation and Inference Methods with Complex Survey Data
Insight Discovery for Decision Tree Models
Jing Shyr
IBM
Jane Chu
IBM
Weicai Zhong
IBM
The decision tree model is a popular data mining tool in predictive analytics. The goal of building it in most applications is for prediction only. The question of identifying which leaf nodes have different target distributions from the root node remains unanswered. Such a missing part should provide further insights and understanding into the predictive structure of the data while it is often overlooked by a tree model user. It might be possible to find some of such leaf nodes in checking the tree diagram when the number of leaf nodes is small or the target distributions between leaf nodes and the root node are very different. However, it becomes more challenging or even impossible when there exist hundreds of leaf nodes or the target distributions between leaf nodes and the root node are not that different. In this paper we propose a systematic and efficient system to identify these leaf nodes based on several tests and present the results in an intuitive way with graphs and texts so it is easy for a tree model user to discover insights. In addition, all tests are based on already computed statistics in the leaf nodes, therefore there is little extra computational cost.