Random Forests for Nonhomogeneous Poisson Processes with Excess Zeros

• Activity Number: 262
• Type: Contributed
• Date/Time: Monday, August 10, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Denis Larocque* and Walid Mathlouthi and Marc Fredette
• Companies: HEC Montreal and HEC montreal and HEC Montreal
• Keywords: Hurdle Model ; Zero-altered Poisson (ZAP) ; Zero-inflated Poisson (ZIP) ; Nonhomogeneous Poisson Process ; Tree-based Method ; Random Forests

Abstract

We propose a method to build trees and forests for a non-homogeneous Poisson response with excess zeros, based on two forests. The first one is used to estimate the probability of having a zero. The second forest is used to estimate the Poisson parameter using trees built with a splitting criterion derived from the zero truncated non-homogeneous Poisson likelihood. Simulation studies show that the proposed method performs well in hurdle (zero-altered) and zero-inflated settings.