Student Seminar Series - January 25, 2006
University of Minnesota
School of Statistics
College of Liberal Arts

Locally Adaptive Combining of Regression Methods


Kejia Shan


Wednesday, January 25, 2006
9:00 AM, B60 Ford Hall
Minneapolis, East Bank Campus

Refreshments at 8:30 AM
300 Ford Hall


Abstract

Various approaches of model/procedure combining have been developed, including Bayesian Model Averaging (BMA) and Adaptive Regression by Mixing (ARM), to aggregate the strengths of candidate procedures. In the current literature, optimal global weights are usually considered. However, in many applications, the candidate procedures may have dramatically different local relative performances. For instance, model 1 may work much better than all other models in one region, but it performs poorly in other regions. In such a case, allowing the weights of the procedures to depend on the covariates can substantially improve the accuracy of the final combined estimator. In this paper, we propose a locally adaptive combining method and examine its theoretical properties. Numerical examples are also provided, which demonstrates the advantage of our method compared to global combining.