I'm writing a custom implementation of PLSA (probabilsitic latent semantical analysis) in SAS/IML with an EM algorithm, but I need to manage a large dataset with circa 17000 words and 250000 documents (a 10% cut dataset for the training job is also too much for IML), but this matrix is very sparse.
So when I calculate the product between word-prob matrix and occurrence matrix, with a non sparse matrix "object", IML returns a "not enough memory" error.