Group decisions arise in various settings, from mundane everyday tasks such as picking a movie to watch at home, to more involved processes such as that of a hiring committee, and require the interplay among group members, which often leads to compromises. Recommender systems, although primarily designed to cater for individuals, have been extended to support group decision making. However, they are hampered by the lack of models that describe the dynamics in group decisions. In this paper, we present ideas about how to extract more appropriate models for explaining and predicting group decisions, by observing decision outcomes.
People are faced with decisions in their everyday life. The need to make decisions as part of a group may not be that often, but can be considerably more complicated, requiring greater e ort and time from all parties involved. Let us consider two separate instances. Assume that a bunch of friends wants to decide on the perfect vacation spot. Each person individually may have di erent and possibly con icting preferences and views as to what an ideal holiday is. As a group though, they have to decide on a single location, which unavoidably means that not all preferences or constraints can be accommodated for. In a similar setting, members of a research group decide on a restaurant to have their monthly meeting.
Both examples portray the same set of challenges, namely the task of reconciling individual incompatible preferences. However, they may substantially di er in the way the decision making process is carried out. In the former setting, it is reasonable to assume that all group members play equal roles and that the nal decision comes as an \impartial" compromise under mutual understanding. In the latter example, however, this may be far from the truth, as the head of the research group has (usually) stronger in uence and authority than the rest.
While recommender systems are typically designed and used to provide
recommendations to individual users, there have been ways to extend them in order
to o er group recommendations. In literature, there are two main strategies for
designing group recommender systems; see also [
We see a missing link here. Existing strategies are well justi ed given that no information on the underlying decision making process is available. Therefore, they have no option but to follow the \least common denominator" approach, e.g., of averaging or aiming for the least misery in the group. On the other hand, understanding and documenting the dynamics of decision making in groups is not an easy task. It is more likely that we can observe and document the outcome of the decision process, rather than its inner workings. Therefore, the question that we raise is whether we can extract the roles and behaviors that govern group decision making by observing its outcomes.
In what follows, we rst review existing approaches in group recommender systems in Section 2, and outline our approach in Section 3. 2
We brie y survey state-of-the-art approaches in group recommender systems.
Group Pro le. Each member of the group is associated with a pro le; in our
setting, a set of item-rating pairs. Methods following this strategy, construct a
separate pro le for the group containing ratings for every item that a group
member has rated. When there are multiple ratings among the group members
for a single item, the item's rating in the group is computed using an aggregation
function; typically, the average or minimum rating is considered. To provide
group recommendations, this merged pro le is treated as a virtual user, and then
standard user recommendation techniques can be leveraged. A comprehensive
survey and evaluation of di erent strategies can be found in [
We remark that the Least Misery strategy is rather conservative, since, in the
absence of any information about group member roles and behavior, it aims not
to displease any member. On the other hand, when member roles/behavior can
be assumed to some degree, e.g., all members have equal saying, it makes more
sense to average predicted ratings. A more detailed evaluation of techniques
aggregating recommendations, with a focus on the goodness of the ranking (instead
of on the ratings accuracy), is presented in [
Furthermore, regarding the two strategies, we note that the former may be
in a disadvantage due to sparse user pro les, which is typically the case. For
example, in the extreme setting when no item is commonly rated by any pair of
group members, it essentially assigns to the group random pro les with no
aggregation. On the other hand, the latter always accounts for all group members,
aggregating their predicted ratings. Note though, that in some settings e.g., [
Our approach is based on the hypothesis that users in groups play distinct, but unknown, roles that determine the outcome of the decision process. Note that it could be possible that the same person exhibits di erent behavior in di erent groups, e.g., in family versus in the work environment. The objective is then to extract the latent roles of group members, in order to explain the observed decisions, and also provide with better group recommendations.
For such an approach to work, we require the availability of decision making outcomes. As a rst step, we assume that we have collected the rating of items by groups as a whole. So in essence the data from which we can learn include user-item as well as group-item ratings.
The next step is to model the group decision making process. In the simplest case, we assume that the group rating of an item is determined by a linear combination of individual predicted ratings, where the weights are the parameters modeling user behavior. We then apply standard learning techniques to extract the weights per user and per group. These derived model parameters can then be used to predict the expected rating of items to groups so as to facilitate and expedite decision making.
In the short-term, we plan to perform a user study to help us evaluate the
techniques. In particular, we will request people to rate items independently,
and also ask them to form groups and provide ratings as a whole. These ratings
will be the ground truth upon which we will evaluate how well our techniques
can predict group behavior. In addition, we will create synthetic group ratings,
where we will simulate di erent decision making strategies among groups, e.g.,
average, least misery, maximum pleasure [
In the longer term, we plan to learn standard prede ned roles for users, e.g.,
dictator, follower, leader, using classi cation models that explain the speci c
behavior of a person in a group. External factors, such as the personality of
the user, that persists across groups, may also a ect the behavior. Additionally,
we plan to consider building richer user behavior models. To achieve this, we
will assume an interactive group decision making system and delve into the
negotiation process [