Fairness in Sequential Group Recommendations

Abstract

In this modern era, sequential recommendations are increasingly prevalent. Users now expect systems to remember past interactions rather than treating each recommendation round as a separate event. Likewise, more and more applications allow users to gather in groups for various activities, such as dining out or movie nights, leading to an increase in the use of group recommendation systems. However, these systems face challenges in processing complex data such as historical user feedback and complex characteristics of the group members. To address these challenges, the SQUIRREL framework is created for implementing Sequential Recommendations with Reinforcement Learning. This framework employs reinforcement learning methodologies to determine the optimal group recommendation algorithm according to the present conditions of the group members. During each round of recommendations, it assesses member satisfaction and item relevance to choose the algorithm that generates the highest reward. While recommending a set of items for a group it is often seen that some members are neglected over others. This makes the recommended list biased to some users' preferences. This work identifies this bias issue and incorporates two new fairness aware reward functions in the original SQUIRREL using the m-proportionality measure. This approach guarantees fairness in recommendations by striving to incorporate a minimum of \textbf{m} items in the group recommendation list that align with the preferences of each member. By incorporating these new functions, the SQUIRREL framework demonstrated its adaptability to changes for additional variables. Evaluation results on the MovieLens dataset illustrate the effectiveness of these new reward functions.