Group Recommendations with Responsibility Constraints

The expansion of social media has led more people to form groups for specific activities, and, consecutively, group recommender systems have emerged as popular research. In contrast to single recommendations, group recommendations involve a much greater degree of complexity since the systems are responsible for balancing the often conflicting interests of all group members. Due to the impact of recommendations on users’ perceived performance (e.g., movie recommendations) and the often inherently sensitive nature of recommendation tasks (e.g., e-health recommendations), the process by which recommendations are generated should be carefully considered. As a result, it has become increasingly necessary to develop recommendations that adhere to various responsibility constraints. Such responsibility constraints include fairness , which corresponds to a lack of bias, and transparency , which facilitates an understanding of the processes of the system.

Nevertheless, if these constraints are followed, group recommender systems be- come more complex. It is even more challenging if they are to consider a sequence of recommendations rather than each recommendation as a separate process. Intuitively, the system should take into account the historical interactions between itself and the group and adjust its recommendations in accordance with the impact of its previous suggestions. This observation leads to the emergence of a new type of recommender system, called sequential group recommendation systems. However, standard group recommendation approaches are ineffective when applied in a sequential scenario. They often produce recommendations that are not even intended to be fair to all group members, i.e., not all group members are equally satisfied with the recommendations. In practice, when each recommendation process is considered in isolation, there is always going to be a least satisfied member. However, the least satisfied member should not always be the same when the scope of the system encompasses more than one recommendation round. This will result in the fairness constraint being broken since the system is biased against one group member.

As a result of the complex nature of recommender systems, users may be unable to understand the reasoning behind a suggestion. To counter this, many systems provide explanations along with their recommendations in adherence to the transparency constraint. Discussing why not suggesting an item is valuable, especially for system administrators. Explanations to such queries are invaluable feedback for them when they are in the process of calibrating or debugging their system.

Overall, this thesis aims to answer the following Research Questions (RQ). RQ1. How to define sequential group recommendations, and why are they needed? How to de- sign group recommendation methods based on them? This thesis formally defines a sequential group recommender system and what objectives it should observe. Additionally, it proposes three novel group recommendation methods to produce fair sequential group recommendations. RQ2. How to exploit reinforcement learning to select a group recommendation method when the system’s environment changes after each recommendation round? In an extension of the RQ1, this thesis proposes a reinforcement-based model that selects the most appropriate group recommendation method to apply throughout a series of recommendations while aiming for fair recommendations. RQ3. How to design questions and produce explanations for why a set of items did not appear in a recommendation list or at a particular position? This dissertation defines what a Why-not question is, as well as presents a structure for them. Additionally, it proposes a model to generate explanations for these Why-not questions. RQ4. How to incorporate various health-related aspects in group recommendations? It is important to make fair recommendations when dealing with extremely sensitive health-related information. In order to produce as fair a recommendation as possible, this thesis proposes a model that incorporates various health aspects.