It was a seductively simple idea that emerged in the early 1990s – to harness the opinions of millions of people online in an effort to help all of us find more useful and interesting content. And, indeed, in various domains and in various forms, this simple idea proved effective. The PARC Tapestry system (Goldberg et al. 1992) introduced the idea (and terminology) of collaborative filtering and showed how both explicit annotation data and implicit behavioral data could be collected into a queryable database and tapped by users to produce personal filters. Less than two years later, the GroupLens system (Resnick et al. 1994) showed that the collaborative filtering approach could be both distributed across a network and automated. Whereas GroupLens performed automated collaborative filtering to Usenet news messages, the Ringo system at Massachusetts Institute of Technology (MIT) (Shardanand and Maes 1995) did the same for music albums and artists and the Bellcore Video Recommender (Hill et al. 1995) did the same for movies. Each of these systems used similar automation techniques – algorithms that identified other users with similar tastes and combined their ratings together into a personalized, weighted average. This simple “k-nearest neighbor” algorithm proved so effective that it quickly became the gold standard against which all collaborative filtering algorithms were compared.

Systems-oriented exploration. With hindsight, it is now clear that these early collaborative filtering systems were important examples from the first of four overlapping phases of recommender systems advances.