Vector piece-wise regression versus clustering (definition and comparative analysis)

Abstract The problem of vector piece-wise regression is formulated. The case where the domain of definition of a vector response function consists of a number of regions of smoothness is considered. The number of the regions and their boundaries are not known and they should be found by analysing a sample of signal corrupted by noise. The solution may be obtained by combination of a dynamic programming algorithm and a probabilistic estimate. The comparison with the approach based on cluster analysis technique, is considered and some experimental results are presented.