GroupLens: An Open Architecture for Collaborative Filtering of Netnews

Inside the paper

The problem

Netnews readers faced more articles than they could reasonably inspect, and a subject line or broad newsgroup was often a poor guide to personal interest. A single shared judgement also cannot account for readers having different tastes.

The idea

GroupLens collects readers’ ratings and predicts how much an individual may like an article from the judgements of people with similar past tastes. Its open design separates news-reading clients from servers that exchange ratings and calculate predictions.

How it works

• Readers rate articles through their news clients, which can also display predicted scores alongside article information.
• Rating servers collect and distribute the ratings, and use agreement on previously rated articles to estimate similarity between readers.
• A prediction combines other readers’ ratings with weights based on their correlation with the target reader’s past ratings.
• Readers can contribute under pseudonyms, while independent clients and servers can participate through the common architecture.

What it showed

The paper describes pilot tests and the design changes they prompted, including changes to interfaces and the distributed architecture. A larger distributed test was still planned, and the authors explicitly say it was not designed to establish that people preferred the filtered experience.

Read it for

Read the GroupLens section and Figure 2 for the architecture, then the prediction discussion and ongoing experimentation section for the limits of the evidence. The related-work survey can be skimmed before returning to the social implications.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.