Reducing the dimensionality of compositional data sets