We summarize the results of an intensive simulation study carried out to compare the performance of three approaches to archetypal analysis regarded as a fuzzy clustering tool: the original approach, namely that of Cutler and Breiman (Technometrics 36(4):338–347, 1994), the Ding et al. (IEEE Trans Pattern Anal Mach Intell 32(1):45–55, 2010) proposal, and the factorized fuzzy c-means algorithm. The artificial data we use in our experiment are generated from polytopes in low-dimensional Rn spaces (2 ≤ n≤ 7) , and comprise a diversity of cluster contexts. The simulation results show that the original proposal is generally a more accurate method to uncover the cluster structure hidden in the data and to reproduce the data themselves. However, this supremacy, if any, is not clear for the data generated from real life problems, and devoted to unsupervised clustering problems.

--

Archetypal analysis,Fuzzy clustering,Matrix factorization,Simulation

• Mathematics - Natural Sciences