21.2 Usage Scenarios and Classification of Group
Recommenders
There are many circumstances in which adaptation to a group is needed rather than to an individual. Below, we present two scenarios that inspired our own work in this area, discuss the scenarios underlying related work, and provide a classification of group recommenders inspired by differences between the scenarios.
21.2.1 Interactive Television
Interactive television offers the possibility of personalized viewing experiences. For instance, instead of everybody watching the same news program, it could be personalized to the viewer. For me, this could mean adding more stories about the Netherlands (where I come from), China (a country that fascinates me after having spent some holidays there) and football, but removing stories about cricket (a sport I hardly understand) and local crime. Similarly, music programs could be adapted to show music clips that I actually like.
There are two main differences between traditional recommendation as it applies to say PC-based software and the interactive TV scenarios sketched above. Firstly, in contrast to the use of PCs, television viewing is largely a family or social activity. So, instead of adapting the news to an individual viewer, the television would have to adapt it to the group of people sitting in front of it at that time. Secondly, traditional work on recommendation has often concerned recommending one particular thing to the user, so for instance, which movie the user should watch. In the scenarios sketched above, the television needs to adapt a sequence of items (news items, music clips) to the viewer. The combination of recommending to a group and recommending a sequence is very interesting, as it may allow you to keep all individuals in the group satisfied by compensating for items a particular user dislikes with other items in the sequence which they do like.
21.2.2 Ambient Intelligence
Ambient intelligence deals with designing physical environments that are sensitive and responsive to the presence of people. For instance, consider the case of a bookstore where sensors detect the presence of customers identified by some portable device (e.g. a Bluetooth-enabled mobile phone, or a fidelity card equipped with an active RFID tag). In this scenario, there are various sensors distributed among the shelves and sections of the bookstore which are able to detect the presence of individual customers. The bookstore can associate the identification of customers with their profiling information, such as preferences, buying patterns and so on.
With this infrastructure in place, the bookstore can provide customers with a responsive environment that would adapt to maximise their well-being with a view to increasing sales. For instance, the device playing the background music should take into account the preferences of the group of customers within hearing distance. Similarly, LCD displays scattered in the store show recommended books based on the customers nearby, the lights on the shop’s display window (showing new titles) can be rearranged to reflect the preferences and interests of the group of customers watching it, and so on. Clearly, group adaptation is needed, as most physical environments will be used by multiple people at the same time.
21.2.3 Scenarios Underlying Related Work
In this section we discuss the scenarios underlying the best known group recommender systems:
• MUSICFX [15] chooses a radio station for background music in a fitness centre, to suit a group of people working out at a given time. This is similar to the Ambient Intelligence scenario discussed above.
• POLYLENS [17] is a group recommender extension of MOVIELENS. MOVIELENS recommends movies based on an individual’s taste as inferred from ratings and social filtering. POLYLENS allows users to create groups and ask for group recommendations.
• INTRIGUE [2] recommends places to visit for tourist groups taking into account characteristics of subgroups within that group (such as children and the disabled).
• The TRAVEL DECISION FORUM [7] helps a group to agree on the desired attributes of a planned joint holiday. Users indicate their preferences on a set of features (like sport and room facilities). For each feature, the system aggregates the individual preferences, and users interact with embodied conversational agents representing other group members to reach an accepted group preference.
• The COLLABORATIVE ADVISORY TRAVEL SYSTEM (CATS) [16] also helps users to choose a joint holiday. Users consider holiday packages, and critique their features (e.g., ‘like the one shown but with a swimming pool’). Based on these critiques, the system recommends other holidays to them. Users also select holidays they like for other group members to see, and these are annotated with how well they match the preferences of each group member (as induced from their critiques). The individual members’ critiques results in a group preference model, and other holidays are recommended based on this model.
• YU’S TV RECOMMENDER [20] recommends a television program for a group to watch. It bases its recommendation on the individuals’ preferences for program features (such as genre, actors, keywords).
21.2.4 A Classification of Group Recommenders
The scenarios provided above differ on several dimensions, which provide a way to classify group recommender systems:
• Individual preferences are known versus developed over time. In most scenarios, the group recommender starts with individual preferences. In contrast, in CATS, individual preferences develop over time, using a critiquing style approach. Chapter 13 discusses critiquing and its role in group recommendation.
• Recommended items are experienced by the group versus presented as options. In the Interactive TV scenario, the group experiences the news items. In the Ambient Intelligence and MUSICFX scenarios, they experience the music. In contrast, in the other scenarios, they are presented with a list of recommendations. For example, POLYLENS presents a list of movies the group may want to watch.
• The group is passive versus active. In most scenarios, the group does not interact with the way individual preferences are aggregated. However, in the TRAVEL DECISION FORUM and CATS the group negotiates the group model.
• Recommending a single item versus a sequence. In the scenarios of MUSICFX, POLYLENS, and YU’S TV RECOMMENDER it is sufficient to recommend individual items: people normally only see one movie per evening, radio stations can play forever, and YU’S TV RECOMMENDER chooses one TV program only. Similarly, in the TRAVEL DECISION FORUM and CATS users only go on one holiday. In contrast, in our Interactive TV scenario, a sequence of items is recommended, for example making up a complete news broadcast. Similarly, in INTRIGUE, it is quite likely that a tourist group would visit multiple attractions during their trip, so would be interested in a sequence of attractions to visit. Also, in the Ambient Environment scenario it is likely that a user will hear multiple songs, or see multiple items on in-store displays.
In this chapter, we will focus on the case where individual preferences are known, the group directly experiences the items, the group is passive, and a sequence is recommended. Recommending a sequence raises interesting questions regarding sequence order (see Section 21.4) and considering the individuals’ affective state (see Sections 21.5 and 21.6). A passive group with direct experience of the items makes it even more important that the group recommendation is good.
DeCampos et al.’s classification of group recommenders also distinguishes between passive and active groups [4]. In addition, it uses two other dimensions:
• How individual preferences are obtained. They distinguish between contentbased and collaborative filtering. Of the systems mentioned above, POLYLENS is the only one that uses collaborative filtering.
• Whether recommendations or profiles are aggregated. In the first case, recommendations are produced for individuals and then aggregated into a group recommendation. In the second case, individual preferences are aggregated into a group model, and this model is used to produce a group recommendation. They mention INTRIGUE and POLYLENS as aggregating recommendations, while the others aggregate profiles.
These two dimensions are related to how the group recommender is implemented rather than being inherent to the usage scenario. In this chapter, we focus on aggregating profiles, but the same aggregation strategies apply when aggregating recommendations. The material presented in this chapter is independent of how the individual preferences are obtained.
21.3 Aggregation Strategies
The main problem group recommendation needs to solve is how to adapt to the group as a whole based on information about individual users’ likes and dislikes. For instance, suppose the group contains three people: Peter, Jane and Mary. Suppose a system is aware that these three individuals are present and knows their interest in each of a set of items (e.g. music clips or advertisements). Table 21.1 gives example ratings on a scale of 1 (really hate) to 10 (really like). Which items should the system recommend, given time for four items?