A new arXiv paper offers two faster ways to sort huge word-document collections into topic clusters - and neither one is a free lunch.
Spectral co-clustering groups documents and the words they contain into related clusters at the same time, which is useful for surfacing themes in large text collections. The catch is that it normally relies on singular value decomposition (SVD), a matrix operation that gets slow and memory-hungry as datasets grow. The paper tests two shortcuts: one that uses random projections to approximate the SVD, and another that combines a partial SVD with element-wise random sampling. Both cut runtime compared to running full SVD, but how much they help depends heavily on how sparse the underlying data is.
That caveat matters more than it sounds. Most real-world word-document matrices are already sparse - most documents don't use most words - and the sampling-based method loses its advantage exactly there, according to the paper's tests. The random-projection method held up more consistently across the datasets tested, making it the safer default for typical text data.
In other words, "randomized" isn't a single speedup button. Which shortcut helps depends on what your data already looks like, and picking the wrong one buys you little.