The effect of pre-smoothing functional data on cluster analysis

We investigate the possible benefits of pre-smoothing functional data before performing a cluster analysis. A simulation study compares the accuracy of clustering results on the basis of the use of unsmoothed functional data-and two smoothed versions of the data-as the inputs in a clustering algorithm. Smoothing is usually found to produce a more accurate clustering, with the best results arising from a novel James-Stein-type shrinkage adjustment to the standard linear smoother. Two real functional data sets are clustered using the competing methods to illustrate the procedure.