learning probabilistic user profiles: interesting web sites, notifying user of relevant changes to...

12
Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman, D. Billsus, S. Gaffney, S. Hettich, G. Khoo, D. Kim, R. Klefstad, C. Lowe, A. Ludeman, J. Muramatsu, K. Omori, M. Pazzani1, D. Semler, B. Starr, & P. Yap Department of Information and Computer Science University of California, Irvine Adaptive Web – Spring 2009 Asli Yazagan

Upload: tabitha-skinner

Post on 01-Jan-2016

213 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman,

Learning Probabilistic User Profiles: Interesting Web Sites, Notifying

User of Relevant Changes to Web Pages, and Locating Grant

Opportunities

M. Ackerman, D. Billsus, S. Gaffney, S. Hettich, G. Khoo, D. Kim, R. Klefstad, C.Lowe, A. Ludeman, J. Muramatsu, K. Omori, M. Pazzani1, D. Semler, B. Starr, &P. YapDepartment of Information and Computer ScienceUniversity of California, Irvine

Adaptive Web – Spring 2009Asli Yazagan

Page 2: Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman,

Introduction

3 agents that helps user to find interesting information

Syskill & Webert Long-term information seeking goals

DICA Monitor user-specified web pages and notify the user only for significant

changes.

GrantLearner Notifies an individual of new research grant opportunities that meet the

learned profile of the individual’s research interests.

Page 3: Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman,

Syskill & Webert

• A user-given topic name and an index page

• Users feedbacks to identify user interest

• Used profile to calculate the probability that

any webpage is interesting to the user.

• Represent interest as keyword vectors.

• Naive Bayesian is used to revise the profile

Page 4: Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman,

Experiment & Result

Question: How accuracy would be learned user’s

preferences when data from several topics are

combined? Polled pages from 5 topics rated by a single user and create a

collection over 450 documents.

User rated % 69.5 of these documents as cold.

Result: Asking user to provide a topic and learning a separate profile per topic is essential for this system.

Problem : words that are informative in a domain are irrelevant for others.

Page 5: Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman,

Do-I-Care Creates and maintains a web page for each topic and report its findings.

Other users using the system can monitor such a web pages to be notified

when another user’s agent has found an interesting change on the web.

Questions:

1. When should a user revisit a known site for new information ?

2. How does a user share new information with others who may be

interested ?

 

Page 6: Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman,

DICA Syskill&Webert

Extracts features from the difference between the current and earlier version of a web page.

Extracts features from a web page.

• User defined target pages.

• Identify changes

• Decide it is interesting or not

• Notify user

• Accept relevance feedback

• Facilitate information sharing

Do-I-Care

Page 7: Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman,

Grant Learner System learns to distinguish between interesting and

uninteresting funding opportunities based on user’s ratings of the descriptions.

Page 8: Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman,

Research Work

Goal: to reduce overall amount of effort required by a user to get useful results from the agent.

Idea: to find additional information to determine which word should be use for features in the profile?

1. Asking to user2. Using lexical knowledge [WordNet]

Page 9: Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman,

Experiment

Page 10: Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman,

Using Lexical Information

Page 11: Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman,

Future Directions

Changes Syskill& Webert interface that allows users to share their profiles with other users.

People can publish their profiles and they can be used by other people

interested in someone’s area of expertise.

Similar approach might be used in education settings.

Instructors train a profile based on their judgments about relevance of items to a class they are teaching. These profiles could be given to students, who obtain an “automated information guide” for their information needs with respect to the class they are taking.

Page 12: Learning Probabilistic User Profiles: Interesting Web Sites, Notifying User of Relevant Changes to Web Pages, and Locating Grant Opportunities M. Ackerman,

Questions

?