Fast Updating Truncated SVD for Representation Learning with Sparse Matrices