21.1 Introduction
Most work on recommender systems to date focuses on recommending items to individual users. For instance, they may select a book for a particular user to read based on a model of that user’s preferences in the past. The challenge recommender system designers traditionally faced is how to decide what would be optimal for an individual user. A lot of progress has been made on this, as evidenced by other chapters in this handbook (e.g. Chapters 2,3, 4,5 and 6).
In this chapter, we go one-step further. There are many situations when it would be good if we could recommend to a group of users rather than to an individual. For instance, a recommender system may select television programmes for a group to view or a sequence of songs to listen to, based on models of all group members. Recommending to groups is even more complicated than recommending to individuals.
Assuming that we know perfectly what is good for individual users, the issue arises how to combine individual user models. In this chapter, we will discuss how group recommendation works, what its problems are, and what advances have been made.
Interestingly, we will show that group recommendation techniques have many uses as well when recommending to individuals. So, even if you are developing recommender systems aimed at individual users you may still want to read on (perhaps reading Section 21.7 first will convince you).
This chapter focusses on deciding what to recommend to a group, in particular how to aggregate individual user models. There are other issues to consider when building a group recommender system which are outside the scope of this chapter.
In particular:
• How to acquire information about individual users’ preferences. The usual recommender techniques can be used (such as explicit ratings and collaborativeand content-based filtering, see other handbook chapters). There is a complication in that it is difficult to infer an individual’s preferences when a group uses the system, but inferences can be made during individual use combined with a probabilistic model when using it in company. An additional complication is that an individual’s ratings may depend on the group they are in. For instance, a teenager may be very happy to watch a programme with his younger siblings, but may not want to see it when with his friends.
• How will the system know who is present? Different solutions exist, such as users explicitly logging in, probabilistic mechanisms using the time of day to predict who is present, the use of tokens and tags, etc [10].
• How to present and explain group recommendations? As seen in this handbook’s chapter on explanations, there are already many considerations when presenting and explaining individual recommendations. The case of group recommendations is even more difficult. More discussion on explaining group recommendations is provided in [8] and under Challenges in our final section.
• How to help users to settle on a final decision? In some group recommenders, users are given group recommendations, and based on these recommendations negotiate what to do. In other group recommenders this is not an issue (see Section 21.2.3 on the difference between passive and active groups). An overview of how users’ decisions can be aided is provided in [8].
The next section highlights usage scenarios of group recommenders, and provides a classification of group recommenders inspired by differences between the scenarios. Section 21.3 discusses strategies for aggregating models of individual users to allow for group recommendation, what strategies have been used in existing systems, and what we have learned from our experiments in this area. Section 21.4 deals with the issue of order when we want to recommend a sequence of items.
Section 21.5 provides an introduction into the modelling of affective state, including how an individual’s affective state can be influenced by the affective states of other group members. Section 21.6 explores how such a model of affective state can be used to build more sophisticated aggregation strategies. Section 21.7 shows how group modelling and group recommendation techniques can be used when recommending to an individual user. Section 21.8 concludes this chapter and discusses future challenges.