Abstract
Group recommender systems (GRSs) have been developed to support group decision mak-ing processes with recommendations expected to satisfy a group of people, not just a single person. Most of the previous research has assumed that users’ preferences are stable and independent, and hence systems solely based on individual preferences, in one shot, can predict a group choice. Unfortunately, in practice, what users choose in a group does not fully match their personal taste as the iterative interaction between group members drives them to adapt their behavior and choices to the group setting. Thus, designing effective GRSs necessitates the dynamic support for the group discussion and the decision-making process.
In this thesis, we aim at leveraging interactive recommendation approaches to facilitate group members in discussing and settling on a final group choice. We particularly focus on: (i) the analysis of how people make decisions in groups, in order to build more useful GRSs; (ii) the development of a GRS that employs and integrates group recommendations into the discussion and decision making stage; (iii) an interactive group model that considers preference knowledge elicited before and during group interaction (i.e., respectively the long-term and session-based preferences); (iv) a generic simulation procedure that simulates social impacts on users’ behavior in order to explore the appropriate combination of long-term and session-based preferences in the alternative group scenarios; and finally (v) a novel group discussion simulation process that models agents’ conflict resolution styles in order to investigate their effect on the outcome of the decision making process supported by the GRS.
We have shown in an exploratory user study that the usability score of our system is greater than a standard benchmark and the proposed group model is able to enhance the perceived recommendation quality. The results of various simulation experiments further prove the efficacy of the proposed model in capturing the changes of the users’ preferences, measured in terms of system ranking performance, and indicate that the optimal combination of long-term and the session-based preferences depends on the specific group scenario. Finally, the evidence of our conflict simulation analysis has shed light on how conflict resolution styles are interlinked with other group factors and how they can influence the group recommendation performance, measured by observing the average individual’s loss in utility for selecting the collective choice rather than the personal one, and the difference in utility obtained by the users with the highest and lowest utility in the group.