Iterative Voting under Uncertainty for Group Recommender Systems (Research Abstract)

Group Recommendation Systems (GRS’s) assist groups when trying to reach a joint decision. I use probabilistic data and apply voting theory to GRS’s in order to minimize user interaction and output an approximate or definite “winner item”. Research Question and Contribution People are sometimes required to reach a joint decision. For example, family members who search online for a TV show to watch together or some acceptance committee who needs to jointly choose which applicants to accept. Recommendation Systems and Social Choice are two domains that address these problems. A group recommendation system (GRS) typically provides recommendations for a group of users considering their specific tastes. In the social choice domain, research in voting theory deals with the similar problem of finding a winner item based on voters preferences (Konzak and Lang 2005). In both domains, it is required to interact with the users to obtain item preferences. Traditionally, it is assumed that an entire list of preferences is required in order to reach a joint decision. In practice, spares rating scenarios are common. Users may wish to state preferences only as necessary, particularly in cases of many available options. Also, bandwidth and communication costs may make it impossible to send all preferences of multiple voters. I therefore aim to keep user interaction at its necessary minimum. Research on GRS primarily focuses on outputting accurate recommendation (Garcia et.al. 2012). Aspects such as group decision making or minimizing communication received little attention. I propose to apply voting theory to GRS in order to determine a “winning item”, an item that certainly suits the group. Copyright © 2012, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. My research addresses practical vote elicitation by modeling a GRS that uses voting rules to support group decision, minimizes user interaction, and facilitates uncertainty by creating and updating probability distributions. I assume the users preferences are unknown in advance, but can be acquired during the process, i.e., a user queried for her preferences is able to submit them. There are four general challenges which motivate my work and contribute to the Group Recommender and the Social Choice research fields: The first is group decision; most existing GRS’s output accurate recommendations to satisfy the group preferences but ignore the group joint decision making process. I believe that an important added value to a GRS would be an interactive process that helps the group to reach a final satisfying joint decision with minimal interactions. I propose to accomplish this by employing voting rules based on methods from the social choice domain. The advantage of voting rules lays in the fact that the rules guarantee a definite winning item. Thus I do not output a recommendation in the traditional sense, i.e. an item with some winning probability and some error margin, but rather output a definite winner, i.e., an item which certainly fits the group's preferences. The second challenge is uncertainty; most previous studies do not consider probabilistic knowledge of the distribution of the users’ preferences for the items. I propose to use probabilistic knowledge to select the next user-item query pair, i.e., to query a certain user for his rating for a specific item. Probabilistic knowledge may not be a common assumption, but is important when attempting to reduce communication. I present an algorithm for estimating probabilistic knowledge so as to demonstrate the feasibility of the assumption. The third challenge is communication cost-sensitive elicitation of preferences; I aim to minimize the number of user-item queries and thus reduce communication costs. The last challenge is repeated system usage; a tradeoff exists between eliciting preferences that will lead to a recommendation of good quality now vs. repeated future recommendations. I Seventeenth AAAI/SIGART Doctoral Consortium