Clustering small datasets in high-dimension by random projection