Got it. Maybe it would be enough to do the projection into PCA space and back project only a very very few individual vectors (cluster centroids and special docs). As to density we can also do with only the top few terms so we could drastically compress the dense vector. Still that would require calculating the inverse of V I think.
For now we'll use the original input vectors when inspecting terms. On Sep 10, 2012, at 11:37 AM, Ted Dunning <[email protected]> wrote: Danger Will Robinson. This will make your vectors dense. That can be disastrous. Even if you don't try to *store* these dense vectors, merely moving them around memory can be prohibitive. On Mon, Sep 10, 2012 at 11:20 AM, Pat Ferrel <[email protected]> wrote: > • Project the vectors back into the original term space. Here the > assumption is that the back projected vectors will be "cleaner" in some > sense.
