RecSys2007 October 22, 2007Posted by Andre Vellino in Collaborative filtering, Data Mining, Information retrieval, Recommender service.
I like small topic-specific conferences like RecSys2007! You get to know a few people, for one thing. But also, there isn’t anyone there, typically, who doesn’t have some very good reason for being there, so you’re almost guaranteed to bump into someone interesting.
I didn’t realize that RecSys2007 is, in fact, the first fully-fledged conference for recommender systems. This subset of the Information Retrieval community has been lingering in the world of workshops in larger conferences and occasional summer-schools, but with an attendance of ~120 participants with from dozens of countries, I’d say the field has come of age.
The KeyNote by Google principle scientist Krishna Bharat made a believable case for the claim that Google News is doing it’s part to increase democracy and free speech by providing a (relatively) unbiased, personalized news aggregation service. I’ve had my share of ambivalence about Google News in the past, perhaps in part because I didn’t trust its way of categorizing clusters of news items nor its relevance measures. But I’m going to give it another shot now.
A few interesting tidbits from Bharat’s talk:
According to Reporters Without Borders, Canada is ranked 5th in world on Freedom of speech. I don’t know what metric RWB uses but presumably it’s detailed in their report. The US ranks 48th and even Switzerland is way down the list as well.
Google News now gives a cultural bias to what counts as “World News” and other general categories. So for example, what counts as “World News” is different in Delhi than it is in New York, primarily because what happens in, for example Pakistan is of greater interest in India than it is in the US. Bharat needed some convincing that what counts as “The World” is geo-context-dependant.
Bharat had good things to say about DayLife and Findory which also provide personalized news aggregation services. Sadly, though, Findory is closing shop on November 1st. So much for increasing the diversity of technology in personalization!
I can’t comment on all the papers in the program, but there ware a couple of interesting trends. One is the number of papers that had something to do with (a) protecting user privacy of ratings used in CF systems and (b) protecting the integrity of legitimate user-ratings against “spam-attacks” by bogus users who want to boost the ratings of their items.
What this tells me is that, increasingly, the focus in Recommender Systems is on second-order issues (trust / privacy) not on accuracy or quality. This is a sign of the technology’s maturity, IMO.