From: ted.dunning@gmail.com   
      
   Giesen2@gmail.com wrote:   
   > If I have X items, how can I estimate the number of user ratings that I   
   > will need to start providing good ratings. I can assume that each user   
   > will enter approximately 20 ratings and that 10 item-item overlap is   
   > necessary to qualify as a "good" rating.   
   >   
   > Has anyone ever done something like this before? Any ideas?   
   >   
      
   I have done a bunch of this sort of thing.  My conclusions have always   
   been that explicit elicitation is very difficult to get right and thus   
   prefer implicit observation of behavior (i.e. to predict behavior,   
   don't ask questions, just watch the behavior).  The biggest systems I   
   did were the recommendation engine for MusicMatch and several fraud   
   detections systems after that.   
      
   The minimum amount of data that you need will vary hugely depending on   
   your domain and how your system works.  I recommend that you determine   
   accuracy by testing with live users.  Presumably you can divert a small   
   fraction of your traffic for testing purposes.  Once you can do that,   
   then drawing a graph of accuracy versus number of ratings is pretty   
   simple.   
      
   For reference, playing about 5-10 songs for the Musicmatch engine   
   results in accuracies in the 50%+ range (and often in the 80%+ range).   
   A standard demo was to have a person name one artist that they liked   
   and have the system give recommendations.  These were generally 80-100%   
   acceptable in the top 10-20 and 50% good in the top 50.   
      
   [ comp.ai is moderated ... your article may take a while to appear. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)