Abstract:
|
In functional linear regression, one classic approach is to perform functional principal component analysis (FPCA) on the functional predictor first and then use the first several leading principle component (FPC) scores to predict the response variable Y. However, the prediction power of those FPCs not necessarily coincides with the amount of variation they account for in the functional predictor. In this paper, we propose a supervised version of FPCA, which can be considered as a generalization of the classic FPCA. It can automatically detect those FPCs which are associated with the response variable. We compare our method with classic FPCA on real data sets and three carefully designed simulation studies.
|