Reflection on “GroupLens: An Open Architecture for Collaborative Filtering of Netnews”
Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P. and Riedl, J., 1994, October. Grouplens: An open architecture for collaborative filtering of netnews. In Proceedings of the 1994 ACM conference on Computer supported cooperative work (pp. 175–186).Vancouver
Summary:
People read news articles and reacted to them but those reactions were not used. In 1994, this paper put forward a system for mining this hidden resource. Collaborative filtering, centered around shared evaluations, helps people make choices based on the opinions of others. GroupLens was a distributed system created here based on the concept of collaborative filtering of netnews, to help people find articles they might like in a pool of articles. The architecture included newsreader clients and news rating servers called Better Bit Bureau (BBB). Newsreader clients helped the user to rate articles after they read them and display the calculated predicted scores. Rating servers (BBB) gathered ratings and made predictions. They were based on the heuristic that people who agreed in the past are likely to agree again. The use of pseudonyms ensured the privacy of the users not affecting the score prediction.
The key feature of this architecture was that it was open: the authors called readers to work on top of the system proposed and inter-operate improved software for clients and servers with the backbone components of the system created in this paper.
Reflection:
GroupLens seems like one of the first recommendation systems based on collaborative filtering design. The notion of people with similar tastes probably liking related things seems simple but has many applications. While reading this paper, I could find similarities between the GroupLens and the present-day recommendation systems on sites like amazon — telling us what to buy, Netflix — suggesting to us what to watch, Spotify- recommending us what to listen to.
GroupLens was built for netnews and had different topics’ articles with threads in separate newsgroups. Comparing it with modern-day lookalikes, I think this design is like GoodReads, Medium, and even subreddits we have on Reddit. Hear me out. Reddit has subreddits on a particular topic and they contain posts as articles that can become a thread if people talk in the comments. People with similar interests can “rate” (upvote/downvote) an “article” (post) and we see the relevant posts on top based on them in the subreddit.
The earlier rating system on GoodReads analyzed which books people might like, based on books they have liked in the past, and books that people with similar tastes have liked. This seemed a lot like the implementation and heuristic used in GroupLens. The “Users buying this also frequently bought” section on Amazon is similar to the rating score prediction of the news article in GroupLens. The significance of this paper is high seeing all the examples in the modern internet era inspired by this system.
Other heuristics like the length of reading the article etc. are in some ways more relevant now than the rating. For example, YouTube suggests what to watch based on how long we and people with similar watch history have watched a related video or creator. The authors investigated techniques for correlating past behavior based on multivariate regression and reinforcement learning concepts. It surprised me to see that because I assumed we have started using those concepts in the past few years. Looks like they existed then but were not as prevalent as they are now, thanks to cheaper large-scale systems and data.
Learning about piggybacked prototypes from our classmate Emily’s presentation on Tuesday, GroupLens was a piggybacked prototype that was built upon Usenet. I also found the concept of killfiles and the use of low ratings relevant to our research project on triggers. The users who strongly dislike a topic can use a mechanism like kill files and low ratings to make the recommendation system learn that the user does not like those topics, so stop suggesting them.
The paper described terms like kill files well, but initially, it was hard for me to understand what netnews and newsgroups meant in the paper because those terms are not used anymore. It was great to learn that the architecture was open, that the concept of open-source existed in the 90s! Also, a tiny detail I enjoyed was that the authors used “she” pronouns instead of “he” when talking about the “user”. :)
They had a section about ongoing research and I got curious to check where GroupLens was right now. I found that the MovieLens recommender system emerged from the GroupLens research lab at the University of Minnesota! It must have emerged from this paper’s architecture on the GroupLens system.
This reflection was written as part of CS5754: Social Computing class under Prof. Sang Won Lee in Fall 2021.
