Text

theses

A group is a collection of humans. Members within a group often share certain characteristics, interests and preferences, along with their individual differences. Such collections of members lead to interesting collective behavior. In this work, we analyze and model the behavior of groups when part of three real-world applications: group recommendation, personalized search and reference group identification. Group recommendation is a variation of the classical problem of recommending items, but where the client is a group rather than an individual. We are interested in the setting where individuals, part of the same group or not, interact regularly with the recommender system. There are two challenges in group recommendation in this setting: 1) historical information about member and group is often missing; 2) members' presence when they ask for a recommendation may be different at different times. We formulate this problem as a group multi-armed bandit problem and design policies for two types of group feedback. We develop a demo system to collect member and group feedback on recommendations to group events and observe the existence of member influence when the group wants to reach consensus. Personalized search results rely heavily on individuals' search and click history. However, a large portion of queries submitted by users each day is new. It is hard to improve search relevance on these queries. We analyzed queries and clicks at group level and observed that individuals' click preferences align well with groups' preferences. With this in mind, we propose cohort models that model each user through groups of users who are similar in one or more dimensions, and facilitate personalized search through cohort's search intent and click preference. Experiments show that cohort models can achieve significant improvement on search relevance, particularly when personal historical data is insufficient. A good way to assess a person is to look at her reference group. A person is considered to be equal or near equal to people in her reference group. We study the problem of finding a group of comparable people for any given researcher, so that we can better represent and understand the researcher in query. To do so, we build researchers' research trajectory with year, publication and venue. Then, we use a trajectory matching algorithm to determine how similar they are and identify relevant candidates. Our algorithm can be easily modified to find more senior researchers whose early stage of their career is comparable to a given junior researcher. We also provide a map-reduce version of our matching algorithm to make it scale well with data.

ETD_6589

Ph.D.

Includes bibliographical references

by Jinyun Yan

doi:10.7282/T3X63PX2

ETD doctoral

The author owns the copyright to this work.