A0243
Title: Clusterwise nonlinear regression with Gaussian processes methods
Authors: Bo Wang - University of Leicester (United Kingdom) [presenting]
Abstract: Clusterwise regression, also referred to as regression clustering, is a technique that combines regression analysis and cluster analysis to discover relationships within data where more than one relationship exists between response variables and explanatory variables. It aims to estimate the different relationships and partition the data points simultaneously. This problem was first introduced by Spath in 1979, and an abundance of further developments have been studied since then. However, almost all clusterwise regression models and algorithms in the literature are based on linear regression, and the nonlinear regressions are limited to the cases where a family of nonlinear functions are assumed and only the unknown parameters are to be determined, such as polynomial functions, Fourier basis functions. We consider clusterwise nonlinear regression problems based on Gaussian process regression without assuming the form of candidate nonlinear functions. A K-means-like clustering algorithm is proposed, and numerical examples demonstrate its effectiveness. The method is also extended to functional nonlinear regression with scalar response and functional and scalar predictors.
