Reading a Balanced Diet: Foraging in Information Communities
IIS- 0102229
Principal Investigator
Joseph A. Konstan
Computer Science and Engineering Unversity of Minnesota
200 Union Street SE 4-192 EE/CS
Minneapolis MN
55455 Ph: 612.625.1831
Fax: 612.625.0572 konstan@cs.umn.edu
http://www-users.cs.umn.edu/~konstan/
Co-PI
JohnT. Riedl
Computer Science and Engineering Unversity of Minnesota
200 Union Street SE 4-192 EE/CS
Minneapolis MN
55455 Ph.: 612.624.7372
Fax: 612.624.7372 riedl@cs.umn.edu
http://www.cs.umn.edu/faculty/riedl.html
Keywords
Recommender systems, collaborative filtering
, content recommendation, information
filtering
Project Summary
This project explores human behavior when faced with
excess information and evaluates interfaces and systems for helping users make
good use of their information-reading time. The methodology is primarily experimental;
as with our previous work on recommender systems, this project will implement
prototype systems and evaluate their effect on and effectiveness for users.
The project addresses five research questions in three key areas: How do people
forage in information abundance? What interfaces best support balanced information
diets? What is the user experience with balance-aware recommenders? How can
a recommender system support role awareness? And, how does information abundance
affect community? We are addressing these questions in the context of a recommender
system for research articles, combining implicit and explicit user ratings of
content with content analysis to explore mechanisms for discovering and reacting
to information overlap and balance.
Publications and Products
A.M. Rashid, I. Albert, D. Cosley, S.K.
Lam, S. McNee, J.A. Konstan, & J. Riedl, "Getting to Know You: Learning
New User Preferences in Recommender Systems", Proceedings of the 2002 International
Conference on Intelligent User Interfaces, January, (2002), p. 127.
Sean M. McNee, Istvan Albert, Dan Cosley, Prateep Gopalkrishnan, Shyong K. Lam,
Al Mamunur Rashid, Joseph A. Konstan, John Riedl, "On the recommending
of citations for research papers", Proceedings of the 2002 ACM conference
on Computer supported cooperative work, vol. , (2002), p. 116.
Dan Cosley, Shyong K. Lam, Istvan Albert, Joseph A. Konstan, John Riedl, "Is
seeing believing?: how recommender system interfaces affect users' opinions",
Proceedings of the conference on Human factors in computing systems, (2003),
p. 585.
Sean M. McNee, Shyong K. Lam, Catherine Guetzlaff, Joseph A. Konstan, and John
Riedl, "Confidence Displays and Training in Recommender Systems",
Proceedings of INTERACT2003, vol. , (2003), Accepted
B. Miller, J. Riedl, and J. Konstan, "GroupLens
for Usenet: Experiences in Applying Collaborative Filtering to a Social Information
System" , bibl. Springer-Verlag, (2002). Book Accepted
of Collection: C. Leug and D. Fisher, "From Usenet to CoWebs: Interacting
with Social Information Spaces"
John Riedl and Joseph Konstan, "Word of Mouse: The Marketing Power of Collaborative
Filtering" , bibl. Warner Books, Inc., (2002).
Project Impact
Successful demonstration of variety of research paper
recommendation algorithms (through a field experiment). Successful identification
of user satisfaction for collections of recommendations (as distinct from individual
recommendations). Initial results suggesting the importance of qualitative evaluation
of recommendation algorithms and variants.
Goals, Objectives and Targeted Activities
We have been developing a major new research Web
site for other researchers to use in finding computer science research papers
they are interested in. This Web site, tentatively called TechLens, will include
the ability to create groups of researcher papers that are related to the interests
of a researcher, and to get recommendations of other papers that might also be
interesting. We have completed preliminary designs of the database design, user
interface, and several recommendation algorithms for TechLens.
Specific targeted activities include: (1) a thorough study of the content,
collaborative, and hybrid algorithms appropriate for recommendation from a large
pool of overlapping articles, (2) gathering and analysis of user data on reactions
to recommendations with high overlap relative to either other simultaneous recommendations
or prior consumed content, (3) interface design and development of an article
recommender, (4) deployment and usage analysis of an article recommender, and
(5) gathering enough baseline data to proceed with more specific studies in
the following year.
Area Background
Broadly speaking, recommender systems are tools that
help people identify worthwhile content from a larger quantity of total content.
Historically, there have been many technologies deployed to this task, including
query technologies (e.g., information retrieval and filtering), and machine learning
technologies. The specific area we focus in, however, is collaborative filtering--the
use of the experiences and opinions of a community of users to create personal
recommendations for each individual. One advantage of such techniques, compared
with query-based ones, is that collaborative filtering can judge quality as well
as topic relevance. In the 1990's, this advantage became important as the scale
of the web led to many queries returning hundreds of thousands of low-quality,
but apparently relevant, matches.
Collaborative filtering systems have been quite successful in both research
and commerce. (We invite you to try a non-commercial recommender system we run,
called MovieLens, at http://www.movielens.org.) Looking into the future, however,
we are concerned that these systems will soon be outpaced by the explosion of
new content. In particular, we see that large quantities of high-quality, highly-relevant
material are being published. Of course, this is good. But we need tools to
help users assemble a "balanced diet" of information so they aren't
forced to read through 20 articles on one topic before they see an article on
a different topic. Indeed, this is what newspaper and magazine editors have
long done in the non-personalized print medium--assembling the best collection
rather than simply collecting the best individual articles.
Area References
We would invite anyone interested in recommender
systems to start with the March 1997 special issue of Communications of the ACM.
Potential Related Projects
None obvious from short list of projects; interested
in identifying them at the workshop.
Project Websites
http://www.cs.umn.edu/research/GroupLens/index.html
Public products, research datasets, and publications done with the support f this
grant and others are compiled on this website.
Online Software
http://www.movielens.org
MovieLens is a typical CF system that collects movie preferences from users and
then groups users with similar tastes. Based on the movie ratings expressed by
all the users in a group it attempts to predict for each individual their opinion
on movies they have not yet seen.
Online Data
http://www.cs.umn.edu/research/GroupLens/
We currently have two datasets available. The first one consists of 100,000 ratings
for 1682 movies by 943 users. The second one consists of approximately 1 million
ratings for 3900 movies by 6040 users. Before using these datasets, please review
the included readme files for the usage license.
Other Resources
http://www.cs.umn.edu/research/GroupLens/
Two additional data sets are available
on the group website for research use:
1. Compaq Research (formerly DEC Research)
ran the EachMovie movie recommender. When EachMovie was shutdown, the dataset
was released to the public for use in research. MovieLens was originally based
on this dataset. It contains 2811983 ratings entered by 72916 for 1628 different
movies, and it has been used in numerous CF publications.
2. Ken Goldberg from UC Berkeley has also released
a dataset from the Jester Joke Recommender System. This dataset contains 4.1
million continuous ratings (-10.00 to +10.00) of 100 jokes from 73,496 users.