Variable Selection for Model-Based Clustering of Functional Data

• Activity Number: 434
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Kyra Singh* and Tanzy Love and Jacqueline Williams and Jacob Finkelstein and Carl Johnston
• Companies: University of Rochester and University of Rochester and University of Rochester and University of Rochester and University of Rochester
• Keywords: Functional data ; model-based clustering ; variable selection ; functional linear model

Abstract

In studying the health effects of radiation, clustering techniques to identify subpopulations with densely sampled functional data are important for detecting late effects of radiation. However, extraneous variables can mask the true group structure. Utilizing a variable selection technique is particularly important in model-based clustering where there is little or no a priori knowledge of the structure or number of groups within the data. Little work on variable selection methods for model-based clustering has been applied to functional data. We propose a greedy search algorithm to integrate variable selection into the clustering procedure, as in "Variable Selection for Model-Based Clustering" (Raftery and Dean 2006) for functional data. At each step, two models are compared using the BIC. One difficulty in implementing this approach is the lack of software available for constructing multivariate fully functional linear models of functional data represented by splines. We avoid this obstacle by creating a full model using a series of univariate partial regressions with the 'fda' package in R. Our new method successfully finds the most important variables for clustering.