Exactly. In informal experiments I have tested these two systems: a) consider all ratings as positive (whether like or dislike), run boolean recommendations and then filter out explicit matches to negative ratings
b) run boolean recommendations on positive and negative separately and output a score that is a linear combination of the two System (a) was much better than system (b). The reason is that a negative vote is telling you more about what the user likes than about what they don't like. In general, it is also important to use data that the user has the most stake in. With video recommendation, using views that lasted >30seconds instead of all view starts gave much better recommendations. Recommendations are often even lower commitment than a view action and the result is that the quality of the data is lower. I equate user-cares with data-is-good. On Wed, Apr 28, 2010 at 10:55 AM, Sean Owen <sro...@gmail.com> wrote: > I think he's getting at what I was trying to get at with my music example. > > A classical music fan, who knows nothing of rock music, will rate lots > of classical music and tend to not rate other things voluntarily. Some > will be rated high and some low, but they'll all be classical and > fairly related. So you can argue a negative rating often indicates > more positive relationship than none at all -- the fact that the user > even knew about some classical music to rate it negatively is > significant. > > So I would also not treat that as stuff to filter or even the > "opposite" of positive ratings. > > 2010/4/28 Tolga Oral <tolga.o...@gmail.com>: > > Ted, I am not sure if I follow what you mean here: > > > >> This is also very dangerous because negative ratings often correlated > > much > >> more tightly with what people like than with what they don't like. >
