|Friday, February 21|
|CS03 Ensemble Modeling||
Fri, Feb 21, 9:15 AM - 10:45 AM
Using Random Model Tree Ensembles to Study Predictor Interactions (302763)Georgiy Bobashev, RTI International
*Barry Swanson Eggleston, RTI International
Keywords: Ensemble models, Interactions, Tree models, Data Mining
Public health, social science, and marketing research involving exploratory analyses can benefit from assessments of potential interactions between candidate predictors of an outcome. Two primary questions here are what variables should be considered as potential interacting predictors and how does one go about mining a large number of potential interactions? In this presentation, an ensemble model will used to demonstrate an analysis that will address both of these questions. The ensemble model will consist of trees containing complex models in the nodes, where an outcome is modeled as a function of known predictors. In this presentation, it will be shown that such ensembles have the ability to identify mixtures of regressions when the predictors defining the separate regressions are members of a set of candidate predictors. Since these trees can have complex models in the nodes, the partition predictors represent variables that interact with the predictors contained in the node models. Because the partition predictor selection process is data driven, these tree ensemble models are a useful tool for mining a large data set for interactions.