(This is how the recommender handles textual IDs for users and items, by the way.)
On Fri, Jan 21, 2011 at 5:57 AM, Ted Dunning <[email protected]> wrote: > Yes. The assignment of features to locations in a fixed size vector is > done > using hashing rather than a dictionary. With a reasonably large vector > collisions will on average not be too terrible. With smaller vectors or > where we are using massive vocabularies due to feature interactions or > simply to worry less, we can use multiple hashing to assign a single > feature > multiple locations. We can prove that the resulting hashed representation > retains all the information we want and that learning a linear classifier > using the hashed representation should work pretty much as well as a full > representation using the same tricks as the random projection guys use > because, well, the hashed representation *is* a random linear projection. > >
