Variable Priority for Global and Individual Variable Selection

Min Lu, PhD: Tree-based learning methods are widely used for analyzing structured data in medicine, biology, and other applied domains, yet most existing importance measures summarize effects at the population level and can obscure meaningful subject-specific variation. In this talk, I will present the Variable Priority (VarPro) framework and its individual extension, iVarPro, which together provide a model-independent approach to variable assessment using rule-based partitions of the feature space. VarPro identifies variables that drive the conditional mean at the population level, while iVarPro estimates local gradients within rule-defined regions to quantify how small changes in a feature influence predictions for a specific individual. I will illustrate the methodology through simulations and a large-scale survival study of patients undergoing treadmill exercise testing. The talk will also include a brief demonstration of the varPro R software for implementing both global and individual variable importance in regression, classification, and survival analysis.