Dr Duncan Watts, the principal investigator for the Small World Research Project (looking at online social networks) was interviewed by Odyssey on WBEZ back in February as part of an hour-long program about network theory. He said in the course of conversation that his results so far suggest that contrary to popular belief (and my own preconceptions), social networks (at least in his experiment) are pretty egalitarian – the most well-connected people are not as important to the overall network as he thought. I don’t see any links to published results on the Small World site, though – has anyone heard any more about what he found?
Archive forJune 5th, 2003 | back to home
Researchers have found that good information doesn’t always drown out bad in recommender systems. In fact, the research done by the creators of movielens shows if you give (for example) a movie a higher rating than it “deserves” other people will also be inclined to give it a high rating. So “innocent” people will unconsciously “play along” with people trying to influence the system and reinforce their dirty work. Unfortunately for the creators of recommender systems, users will notice when overall a recommender system’s results are poor. The writers of the academic paper (available in full here) suggest one way to avoid this problem would be to hide the rating of a film from users who want to rate it themselves so they aren’t influenced by others’ ratings.
I didn’t find collaborative filtering useful when I did use it, but that was nearly ten years ago when the MIT Media Lab was playing about with what became Firefly. Perhaps if my DVD Recorder was smarter and networked with other such recorders to compare my TV/film preferences with others’ without my needing to enter the details by hand it would have enough data to be able to adequately predict my viewing tastes. Personally I suspect mine are atypical enough that it would be difficult to predict what I would like mathematically. Then again, most people probably think they are unique in this respect!