Ph.D. Defense of Dissertation Announcement

Title: Personalized Search and Recommendation for Health Information Resources

Steven P. Crain
School of Computational Science and Engineering
College of Computing
Georgia Institute of Technology

Date: Tuesday, August 7th, 2012
Time: 12:30 PM - 1:30 PM (EDT)
Location: KACB 1315

Committee:

• Dr. Hongyuan Zha (Advisor, School of Computational Science and Engineering, Georgia Tech)
• Dr. Alexander Gray (School of Computational Science and Engineering, Georgia Tech)

• Dr. Amy Bruckman (School of Interactive Computing, Georgia Tech)
• Dr. Mark Braunstein (School of Interactive Computing, Georgia Tech)
• Dr. Eugene Agichtein (Department of Mathematics and Computer Science, Emory)

Abstract:
People are increasingly using the Internet to fill health-related needs.
They search for information that helps them better manage their health and understand their conditions. They also seek out communities where they can learn about health together and get social support. Users may face several challenges in this context. First, in many cases consumers face a language gap as they look for information that is written in unfamiliar technical language. Second, it can be difficult for a user to find the most relevant, interesting and useful health-related groups. We present solutions to these challenges.

Topic modeling is commonly used to find information when searching for a general set of ideas instead of specific words. Conventional topic models discover a set of unrelated topics that together explain the combinations of words in a collection of documents. We add additional structure that provides relationships between topics corresponding to relationships between consumer and technical medical topics. This allows us to support search for technical information using informal consumer medical questions.

We study the process by which users participate in discussion groups in an online diabetes community, TuDiabetes. We develop novel event history analysis techniques that allow us to identify the the partially observable group characteristics that are correlated with user activity in the groups. This analysis reveals that the current mechanisms provided for finding groups, which uniformly advertise the popular groups to all users, impair the diversity of the groups and limits their value to the community. Accordingly, we develop a system that remediates this problem by providing personalized recommendations of groups to join based on a user's previous interactions with other groups. This increases the variety of groups that are visible to a user and also promotes diversity between groups.

We evaluated the ability of our structured topic models to assist with finding health information using a collection of consumer questions taken from Yahoo! Answers health categories and documents of various technicalities taken from consumer-oriented websites and technical medical documents. Mechanical Turk was used to obtain identify technical documents that were relevant to the consumer questions. Based on this data, we found that the structured topic models performed around twice as well as conventional topic models in terms of normalized discounted cumulative gain. We evaluated the group recommendation system by measuring how well it predicted a sample of the groups actually joined by users. When each test group was mixed with 30 random groups, it was on average ranked fifth, which indicates that the system is likely to produce more useful recommendations than a popularity-based uniform recommendation.