CHI Nederland


1 november 2001

Door:
Geert de Haan

Collaborative Filtering Recommender Systems

door: Joseph A. Konstan, University of Minnesota, USA

Introduction

Joseph Konstan started his presentation explaining the reasoning behind his work: at first, computer users were faced with the problem of finding information. Recently, the arrival of the internet created the problem of information overload. Recommender systems try to do something about it by means of filtering (e.g. email), recommendation (e.g. web sites) and by prediction (e.g. movies).

Information retrieval can be characterised by 1) the information needs are dynamic and 2) the content base is static. Information filtering can be characterised by 1) the information needs are static and 2) the content base is dynamic. Collaborative filtering connects the two in that the information needs go beyond simple keywords or topics and involve things like quality and taste. Presumably, collaborative filtering works better with smaller communities, such as the community that shares ones taste for particular movies.

Collaborative Filtering (CF)

GroupLens (Resnick et al. at CSCW'94) is the published example of a collaborative filtering system, soon followed by other systems. Presently collaborative filtering is everywhere (e.g. amazon.com, wine.com, cdnow.com, MovieLens, etc.)
GroupLens connects communities that do not know each other by collaboratively filtering of Usenet news based on user rating of news items, identification of user types, and prediction of what particular users might be interested in. It ran for 7 weeks with 250 registered users on several newsgroups. It ran successfully since users liked it and used it long after the trial period and it actually changed people's behaviour. In addition, the trial proved that prediction for individuals works better than prediction for the average or the modal user.

CF works on the basis of the rating that people supply in exchange for a prediction or recommendation. One rates a number of movies or news items. These ratings are correlated with those of others, and using (dis-)similarities in the pattern of ratings makes it possible to predict which movies or news items a particular person will appreciate. Technically, finding patterns of correlations among ratings is done by some kind of self-organising neural networks like Kohonen nets. To make the proper recommendation (rather than speed) additionally requires considerable parameter tweaking.

In a broader context, CF is one method to attain the same goals as, for example, information filtering, intelligent agents which are rather 'wrappers', content space navigation like the restaurant recommender (if you like Wimpy, you will like the frietboer), social navigation like using footprints (follow where other went, until everyone does the same), and pull/push collaborative filtering. All these approaches attempt to lessen information overload and all suffer from the self-fulfilling prophecy of progressively filtering away more and more material.

Interfaces and the user experience

Under this heading Konstan discussed several experiments and findings related to the MovieLens, a system which recommends which unseen movies someone will appreciate based the persons likes or dislike about seen movies in relation to those of other people.

challenge: coping with belief.

People may like recommender systems even though they may not work properly. In one experiment the amount of explanation was varied and the results show three bands, a middle band with no explanation, a low band of belief because of the recommendation, and a high band because of the content correlation. Of these, the content correlations worked best.

A second experiment addressed ephemeral (immediate, temporal) needs, such as, being in the mood for a James Bond type of movie. Selection of recommendations is much appreciated, and in this case, the precision and relevance of the recommendations depend very much on the support threshold or the number of ratings, and there is a need for counterexamples.

challenge: the value of information.

In building recommender systems, not based on Some results are: the value of information is higher when it is contentious e.g. when the experts disagree, when the community is small or the information rare. The value of information is lower for popular information and recommendations, such as, best-selling items. Here communities may build very fast.

challenge: recommendations for groups.

PolyLens is a movie recommender system aiming at groups of people because people tend watch films together. In this context, questions arise like, what is a group, how to form one, and questions about privacy issues. In an experiment, people could create groups and join or leave them on the basis of access to shared individual recommendations (read: ratings). Here, results indicate that people liked the system and thought it useful, and that privacy was not an issue or rather: access to other people's recommendations was considered essential.

challenge: meta-recommendation.

Recommendation may not prove useful when they do not take the user's context into account, like the price of the item, or whether a movie is currently playing nearby. Adding such facilities to the recommender system enormously increase both the utility of the recommendations as well as the user's appreciations.

CF under diminished returns

In the world as we know it, people's like for chocolate will decrease with the number of chocolate bars eaten. This creates some challenges for the future of recommender systems. One is, how to deal with the increasing familiarity or expertise of a person. To the initial user, general recommendations will suffice but to satisfy expert users, more and more needs be sifted out.

A second challenge is to address the need for change. Newspapers would not sell very well when they only contained the best articles because many will be about the same subject. As such, there is a need to determine, in addition to the best, the best portfolio. A third and final challenge, which has not been worked on is, how to cope with the changes in taste, such as seasonal changes and changes in the personal taste of people.

About the presenter

Joseph A. Konstan is Associate Professor of Computer Science and Engineering at the University of Minnesota. His research addresses a variety of human-computer interaction issues related to filtering, comprehending, organizing, and automating large and complex data sets. He is probably best known for his work in collaborative filtering (the GroupLens recommender system) and multimedia authoring. In 1996, he co-founded Net Perceptions, Inc., a company that has since developed collaborative filtering systems into a variety of commercial personalization tools.

Dr. Konstan received his Ph.D. from the University of California, Berkeley in 1993; he is an ACM Lecturer, editor of the SIGCHI Bulletin, and a member of the ACM SIGCHI Executive Committee.

A variety of research papers on GroupLens can be found at www.GroupLens.org, and papers on MovieLens, the current research system, can be found at www.movielens.org.

Thanks

SIGCHI.NL wishes to thank the Vrije Universiteit, Amsterdam for providing hospitality for this evening meeting.

author: Geert de Haan