Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Machine learning and social choice theory for personalized and group recommendations

Streviniotis Errikos

Full record

---

Summary

Recommender systems are software tools that provide assistance to individuals who lack experience or knowledge in order to overcome the information overflow problem. In this work, we introduce two recommender systems that employ Bayesian learning and the Kalman Filter algorithm respectively, along with mechanisms derived from Social Choice Theory, in order to learn users preferences and provide efficient personalized recommendations and group recommendations.To facilitate user preference learning, we propose a novel, lightweight preference elicitation process, during which the user is presented with and asked to rate a small number of generic images that are related with the items under recommendation. We then exploit these ratings to guide our approaches to generate beliefs regarding the user's preferences. In order to model the high uncertainty that exists in such settings, our system represents both users and items as multivariate normal distributions.On top of these, we employ several multiwinner voting rules from the social choice literature to the personalized recommendations problem. Specifically, we equip with such mechanisms our recommenders, allowing for effective personalized recommendations while promoting diverse results with respect to several features. We then focus on the group recommendation problem, by extending our approach and employing various preference aggregation mechanisms alongside with a multiwinner voting rule, namely the Reweighted Approval Voting (RAV). The application of multiwinner voting rules for these problems is, to the best of our knowledge, done for the first time in the literature.Thus, in this thesis we tackle both the personalized and group recommendations problems; and we do so focusing on the tourism domain. We conduct a systematic experimental evaluation of our approaches by applying them on a real-world dataset of Points of Interest (POIs) in the popular touristic destination of Agios Nikolaos, Crete, Greece. Interestingly, we study the effectiveness of our approaches when we equip our system with prior knowledge regarding the (average) preferences of specific user types (i.e., tourists belonging in specific age groups), given data we collected via questionnaires from actual tourists visiting the city of Agios Nikolaos.Our experimental results (i) highlight the ability of our systems to successfully produce personalized recommendations that match the specific interests of a single user; (ii) confirm that the employment of prior knowledge regarding the preferences of tourists, based on their demographics, guides our recommender to avoid the cold-start problem; (iii) demonstrate that the use of multiwinner mechanisms allows for diverse recommendations with respect to travel-related features, and increased system performance in the case of limited user-system interactions; and (iv) show that the use of multiwinner mechanisms allows for fair group recommendations with respect to the well-known m-PROPORTIONALITY and m-ENVY-FREENESS metrics. Last but not least, our personalized Bayesian recommendation algorithm is incorporated in a real-world mobile tour-planning application for Agios Nikolaos, Crete.

Available Files

Services