That's a good point, on keeping it for a while, however in our case items that are removed are not relevant, they have been added by mistake or are inappropriate, so basically they don't bring any value. Also, handling 'ghost' items adds additional complexity...
We don't generate recommendations for each user in background... the model is loaded in memory when taste starts up(along with tomcat). Data for the model is being loaded from database. The actual recommendations are retrieved on user's requests and each time user makes an action on items (likes, etc..) he may see different results in recommendations... On Sun, Oct 16, 2011 at 9:25 PM, Ted Dunning <[email protected]> wrote: > On Sun, Oct 16, 2011 at 6:58 PM, Octavian Covalschi < > [email protected]> wrote: > > > ... We have the taste war running and we are trying to accomplish a > > realtime > > functionality by removing/adding user items on the fly. So far adding and > > removing items for particular users has been working fine for us and > > recommendations are being changed on the fly.. though it's still in > beta... > > The last piece is to remove an item from ALL users, it's necessary when > the > > item is removed from the system... > > > Two thoughts on this. > > First, is it really necessary to remove the item at all other than from > recommendation results? After all, the item still defines some aspect of > user similarity. Why not keep it around for a while? > > Secondly, can you do the removal lazily as you start to generate > recommendations for a user or in the background in a scan over all users? >
