Learning to Recommend

Haym Hirsh (Computer Science Department, Rutgers University)

Most "collaborative" filtering systems recommend new artifacts by
comparing your ratings of a collection of artifacts to the ratings of
other people.  Artifacts liked by others who have similar ratings
profiles on these artifacts are then recommended to you.  But what if
you also have further information about each of the artifacts -- can
you improve the recommendation process by using "content-based"
information?  This talk will describe a machine-learning approach to
combining collaborative and content-based filtering for movie
recommendation.  We combine collaborative information -- available as
a collection of 50,000 movie recommendations across over 250 users
collected by an existing collaborative-filtering system -- with
context-based information -- in the form of 27 features for each
movie, such as year, genre, and actors in the movie.  Our experimental
results show that our machine-learning approach successfully exploits
the additional content-based information to improve recommendation
beyond what was achieved using collaborative filtering alone.