GroupLens: predictions from other readers' ratings
On 22 October 1994, at the CSCW conference, Paul Resnick of MIT, John Riedl of the University of Minnesota and three co-authors described GroupLens, an open architecture for collaborative filtering of Usenet news. Readers rate articles; rating servers called Better Bit Bureaus gather the ratings and predict how another reader will rate an article, on the heuristic that people who agreed in the past will agree again.
Why it matters
The prediction draws not on an article's content but on correlations between people's ratings, and it runs as an open service between sites rather than a filter inside one organisation. Shops and feeds later deployed the principle: Amazon's paper of 2003 cites GroupLens as the model of traditional collaborative filtering.
The paper gives scale only for Usenet itself: by the estimate of 24 January 1994 more than 140,000 people had posted in the previous two weeks, with traffic over 100 MB a day. GroupLens itself had been through a pilot of four people at the University of Minnesota (an earlier version with a slightly different scoring function), and before that seven people at a Schlumberger research lab had tried an endorsement mechanism; a larger distributed test between MIT and Minnesota is only planned. Ratings can be entered under pseudonyms. Tapestry, a single-site system, is named as the predecessor. The record does not claim that GroupLens had users beyond the pilots in 1994.