Evaluating Diversity in Sequential Group Recommendations

Abstract

As the complexity of user preferences increases in group activities such as watching movies, more dynamic recommendation systems are needed. The SQUIRREL framework, utilizing Reinforcement Learning (RL), adapts to the changing preferences of group members through sequential group recommendations. This thesis introduces a significant enhancement to SQUIRREL framework by permitting the re-recommendation of items, reflecting more realistic scenarios where users may revisit previously enjoyed content. We implemented and evaluated three distinct reward functions. These functions focus on maximizing group satisfaction, balancing satisfaction with diversity, and maximizing diversity. The goal was to investigate their impact on recommendation diversity and group satisfaction. Experimental results, using the 20M MovieLens dataset, demonstrate that while diversity-focused reward functions enhance recommendation variety, it requires a careful balance with user satisfaction. This is necessary to avoid diminishing the overall user engagement. The study confirms the effectiveness of adaptive reward functions in improving the engagement and satisfaction of group recommender systems, thus highlighting the need for new methods for producing group recommendations that ensure highly relevant and diverse at the same time results for groups. Our research extends the capabilities of the SQUIRREL framework, showcasing how adaptive systems can better meet the complex demands of the group dynamics in recommendation scenarios. The study further emphasizes the significance of continual adaptation and fine-tuning of recommendation algorithms. These approaches effectively address both the individual and collective preferences of users, thus enhancing the overall user experience in group settings.