I have heard arguments both ways for the off-line cooccurrence analysis. Certainly, the user getting the recommendations should get them based on the most recent items.
On Mon, Dec 5, 2011 at 12:34 AM, Daniel Zohar <[email protected]> wrote: > I think it would be extremely good to be able to recommend on items related > to more recent preferences. That is, if the timestamps are provided. > > On Mon, Dec 5, 2011 at 10:27 AM, Ted Dunning <[email protected]> > wrote: > > > The downsampling should have a target size after sampling so that users > > with that many or fewer ratings are not down-sampled at all. > > > > This is easy to do using reservoir sampling or anything similar. You can > > also just keep the first or most recent ratings. Or you can use a > sampler > > biased toward either of those extremes. > > > > On Mon, Dec 5, 2011 at 12:20 AM, Daniel Zohar <[email protected]> > wrote: > > > > > Another issue to take into account, is to try and not down-sample too > > much > > > so users with 1-2 preferences still get decent results. > > > > > >
