Community Lab - Projects

Risks to K-Anonymity in Recommender Systems Datasets

Thu, 13 Jul 2006 19:25:06 -0400

How secure is personal identity in de-identified data sets?

To what extent can users be re-identified between the publicly available datasets. The results show that even when the movie identification from the forums is performed with very simple text analysis algorithms 31% of users can be 1-identified; with hypothetical more sophisticated text analysis, we estimate more than 40% of users could be 1-identified.
How much of a database must be redacted to prevent 1-identification? The results show that more than 80% of the low popularity items must be removed from the database to reduce 1-identification to near zero. (These 80% of the items only represent about 20% of the total ratings in the dataset, because each of the items has few ratings.)
Can a user protect herself from 1-identification through careful choice of the items she mentions in the forums? Here the results are mixed. The simple approach of not mentioning some movies is relatively ineffective: about 30% of the movies a user might like to discuss must be left out of the postings. On the other hand, the more subtle approach or misdirecting by mentioning movies that other users find interesting does work more easily. There is some question about the ethics of this approach, because it will redirect the re-identification to some other user.

Project members:

Title	Authors	Appears In	Publication Date	Date added
You Are What You Say: Privacy Risks of Public Mentions	Frankowski, D., Cosley, D., Sen, S., Terveen, L., Riedl, J.	Proceedings of SIGIR 2006	2006	07.13.06

Social Preferences

Thu, 13 Jul 2006 18:49:18 -0400

An exploration of the use of experimental economics techniques to model user motivation for participation, and to appeal to users in different ways to change participation.

Developed a theoretical model to analyze motives for contribution to online communities. Designed and carried out a survey of 400 MovieLens users to test the assumptions of the model and to develop practical models for classifying users into different motivational subgroups. Have designed (and will carry out) experiments involving eliciting more substantial contributions using cues targeted at users' motives. Goal of work is to test social preference and conformity theories.
Theoretical analysis shows that:

users who want to influence others contribute more ratings than users who only care about their own direct benefits;
if some users are inequality averse, publishing the distribution of ratings will make user ratings converge to the mean of the distribution; and
when there are a large fraction of selfish users, the total amount of rating on MovieLens will be less than the socially optimal amount.

Developed the first empirical economic model of user contribution to an on-line community based on a combination of factors including intrinsic benefit, user effort, and direct and indirect benefits.
Validated the first experimental economics model of user contribution to an online community built on user utility derived from direct and indirect sources.
Demonstrated experimentally the effectiveness of both conformity and inequality aversion in shaping the behavior of users in even a low-interaction, low awareness online community.
A case study in how to conduct experimental economic experiments in an online community using a combination of behavioral and survey data.
Found that both conformity and inequality aversion affect user behavior, but in different ways. Conformity worked as expected, leading lower-contributors to increase their contribution. Inequality aversion steered high-net-benefit users towards higher-effort (more altruistic) activities. Findings are still being analyzed and will be submitted as a paper in the near future.

Project members:

Title	Authors	Appears In	Publication Date	Date added
An Economic Model of User Rating in an Online Recommender System	F. Maxwell Harper, Xin Li, Yan Chen, and Joseph A. Konstan	Proceedings of The 10th International Conference on User Modeling	2005	01.28.06

Tagging Ontology Emergence

Wed, 12 Jul 2006 11:45:30 -0400

Five research questions were investigated in the study of tagging:

How strongly does personal tendency affect personal tagging behavior? Our data show a strong effect from personal tendency.
How strongly does community influence affect personal tagging behavior. Our data show that the tags a person has seen from the community also strongly influence future tagging, though less strongly than personal tendency.
How does the tag selection algorithm influence the evolution of the community's vocabulary? Our data show significant differences among the experimental groups in tagging behavior, though it is hard to know for certain that the tag selection algorithm is the cause of the difference in behavior, given the strong effects of personal and community tendencies.
How does the tag selection algorithm affect users' satisfaction with the system? Our data show strong differences in satisfaction with the different algorithms, though the differences in overall satisfaction are less strong.
Are different tag classes more or less well-suited to various user tasks? Our data show that users definitely prefer factual tags for learning about movies, subjective tags for choosing movies, and personal tags for categorizing movies.

Project members:

Shilad Sen

Title	Authors	Appears In	Publication Date	Date added
tagging, communities, vocabulary, evolution	Sen, S., et al.,	Proceedings of CSCW 06	2006	07.13.06

Making the Social Visible

Wed, 12 Jul 2006 11:26:57 -0400

Exploring how cues showing the presence of other users might elicit more social responses from MovieLens users.

Conducted a study to determine whether social cues elicited different levels of contribution. The study started with a visual re-design of elements of the MovieLens site to change language and other displays to make more evident the social nature of the site. A subsequent study did not find differences between users given the new and old interfaces (which we find to be an interesting negative result).
Some of this work is continuing as we look at cumulative and interaction effects when social cues are used with more genuinely social features.
Found that textual cues about the presence of other users are not sufficient to change users’ social reaction to a recommender site.
Provided evidence that social cueing requires more than simple textual cues.

Project members:

Reinventing Conversation

Sun, 22 Jan 2006 04:33:15 -0500

Discussion in online forums typically consists of threaded messages, presented in a static (typically chronological) view. There is basically just one view of a conversation, and threads are attached to topic- specific forums, regardless of how their content evolves over time. This model doesn't allow for much flexibility for how information is presented to the users. We are investigating methods for creating a dynamically structured conversation space in which to organize threaded conversation. Rather than being part of a static topic that may not accurately represent topic evolution, threads can be associated dynamically with various relevant objects, e.g., a thread discussing participants' favorite French New Wave films may be accessed from the page for any of the movies that were mentioned. This makes it possible for users to encounter and initiate conversations in a variety of contexts. Breaking away from the chronological, segmented view of threads, users will instead be able to see messages posted in any area of the website, at any time, by any user. This new approach allows for a dynamic display of the conversation space, and the capability of presenting those threads which are of most interest to the user.

• Related Projects: MessagePlus

Project members:

Title	Authors	Appears In	Publication Date	Date added
Insert Movie Reference Here: A System to Bridge Conversation and Item-Oriented Web Sites	Drenner, S., Harper, M., Frankowski, D., Riedl, J., and Terveen, L.,	Proceedings of CHI 2006	2006	07.13.06

MessagePlus

Sat, 21 Jan 2006 08:13:20 -0500

Readers frequently encounter messages from online conversations out of their original context. We develop a family of algorithms for presenting parts of the conversation surrounding a focal message; we refer to them as MessagePlus views. In a lab experiment, subjects using a MessagePlus interface were able to complete tasks more quickly than those using full displays of the same threads.

In a lab experiment, users were able to make faster decisions about the relevance of the information in the conversation using the MessagePlus interface than the interface with everything displayed. There was no significant difference in accuracy.
The experimental interface to Slashdot launched in summer 2006.

Stay tuned for results and papers about this work.

Project members:

Self-Maintaining Communities

Fri, 20 Jan 2006 21:22:22 -0500

Most members' contributions to an online community pertain to what Preece calls its purpose. Members post to discussion groups, rate movies and receive recommendations, and read each others' blogs. These contributions are visibly important and constitute the day to day business of the community. However, communities need inputs besides conversation, such as moderation, governance, and the maintenance of databases (members, movies, FAQs, and histories, for instance). These duties usually fall to the owners of the community. Owners don't have to be the only people who can perform maintenance. Web sites could be designed to allow all members to contribute information. By taking advantage of all members' knowledge and effort, member-maintained communities can reduce their dependence on key members while increasing their overall value to everyone and reflecting everyone's desires. But letting members contribute is not a panacea. Most systems have no interface for contributing. Members need to be motivated to do the job and trained to do it well---and perhaps watched also, because some will make mistakes and others may sabotage the group. We believe that social science theories that address why people contribute to groups can be a valuable tool for designing interfaces and increasing people's motivation to contribute.

Project members:

Title	Authors	Appears In	Publication Date	Date added
How Oversight Improves Member-Maintained Communities	Cosley, D., Frankowski, D., Kiesler, S., Terveen, L., Riedl, J.	Proceedings of CHI 2005	2005	07.13.06
SuggestBot: Using Intelligent Task Routing to Help People Find Work in Wikipedia	Cosley, D., Frankowski, D., Terveen, L., and Riedl, J	Intelligent User Interfaces	2007	06.06.07
Using Intelligent Task Routing and Contribution Review to Help Communities Build Artifacts of Lasting Value	Cosley, D., Frankowski, D., Terveen, L., & Riedl, J.,	Proceedings of CHI 2006	2006	07.13.06
Virtual Community Maintenance with a Collaborative Repository	Hansen, Derek, Ackerman, Mark, Resnick, Paul, and Munson, Sean	to appear in Proceedings of ASIST 07		07.03.07

Value of Contributions

Fri, 20 Jan 2006 18:08:32 -0500

Measuring the value of specific contributions users make to a community and using that measurement to influence participation.

Developed algorithms for computing the value of ratings both prospectively and retrospectively.
Designed, conducted, and wrote up an experiment to test whether making the value of a contribution explicit increased contribution and if the beneficiary of the contribution (self-versus other) made a difference. This paper was accepted at the ACM Conference on Computer Supported Cooperative Work.
Designed (and will carry out) an experiment to test the effect of revealing value to the users, and of the effect of the size and type of the beneficiary group, on user contribution.
Our initial experiment showed, as predicted by theory, that individuals contributed more when they were reminded that their contributions were unique. However, a puzzling result was that reminding individuals that their ratings help either themselves or others prompted less rather than more rating.
Developed new algorithms for assessing the value of a specific rating contribution to a community, both retrospectively for past contributions and prospectively for potential ones.
Designed new interfaces to make this information visible to users.
Conducted experiments where we asked subscribers to MovieLens to rate additional movies as part of a campaign. In prior experiments in the series, we attempted to signal both the beneficiary of contributions and their value through persuasive email messages sent to subscribers. In the newest experiments, we modified the user interface to MovieLens so that the beneficiary of a rating and its value are shown as an icon associated with each movie. We developed algorithms to assess the value of a rating, which we defined as improvement in accuracy of predictions about a movie for a particular target group if a subscriber rated it. We showed subscribers the relative value of rating 100 movies they were most likely to have seen for each of four target groups -- the subscriber him or herself, the average movie lens subscriber, someone who liked the movie genres the subscriber liked, and someone who like genres the subscriber did not like.
Subscribers participating in the rating campaign rated approximately 7.3% of the movies shown to them. Their likelihood of rating a movie doubled when they believed that they were helping other subscribers who liked genres they liked and increased by about 50% when they believed they were helping the typical MovieLens subscriber. However, they did not increase the number of rating they made when they believed that doing so would help themselves or would help subscribers who liked genres they did not like.
Expanded upon prior work in 'value of information' analysis to present new and more focused algorithms for assessing the value of a rating contribution.
We have extended our exploration of algorithmic approaches to assessing actual and potential value of contributions to a user and to the community as a whole.
The first experiment found that users were more likely to rate items that were indicated as high value. Users self reported that they would most prefer to rate items that were high value for themselves; in practice, however, they actually preferentially rated items that were marked as being of high value for other users. In other results, users chose to rate items that were of high value to small groups of similar people, in preference both to dissimilar groups, and to the population as a whole. We have not yet completed data analysis on the second experiment.

Project members:

Title	Authors	Appears In	Publication Date	Date added
Group theory for social engineering: Promises and pitfalls	Robert Kraut	Presented at the annual meeting of the Academy of Management	2004	01.28.06
Influence in Ratings-Based Recommender Systems: An Algorithm-Independent Approach	Al Mamunur Rashid, George Karypis, John Riedl	Proceedings of SIAM 2005 Data Mining Conference	2005	01.28.06
Modeling Member Motivation and Participation in Online Communities	Ren, Y., and Kraut, R.	Proceeding, Academy of Management Conference	2006	07.13.06
Motivating Participation by Displaying the Value of Contribution	Al Mamunur Rashid, Kimberly Ling, Regina D Tassone, Paul Resnick, Robert Kraut, John Riedl	Proceedings of ACM CHI 2006 Conference on Human Factors in Computing Systems	2006	01.25.06
Think different: increasing online community participation using uniqueness and group dissimilarity	Pamela J. Ludford, Dan Cosley, Dan Frankowski, and Loren Terveen	Proceedings of the SIGCHI conference on Human factors in computing systems	2004	01.28.06
Using Social Psychology to Motivate Contributions to Online Communities	Gerard Beenen, Kimberly Ling, Xiaoqing Wang, Klarissa Chang, Dan Frankowski, Paul Resnick, Robert E. Kraut	Proceedings of ACM CSCW 2004 Conference on Computer Supported Cooperative Work	2004	01.28.06
Using social psychology to motivate contributions to online communities. Journal of Computer-Mediated Communication	Ling, K., Beenen, G., Ludford, P., Wang, X., Chang, K., Cosley, D., Frankowski, D., Terveen, L., Rashid, A. M., Resnick, P., and Kraut, R	Journal of Computer-Mediated Communication	2005	01.28.06