One alternative to look at for some datasets is nonnegative matrix factorization. It's similar to PCA, but every element of each basis vector has to be positiv

One alternative to look at for some datasets is nonnegative matrix factorization. It's similar to PCA, but every element of each basis vector has to be positive. This tends to make the results a lot easier to interpret.