Although most recommender systems make suggestions for individual users, in many circumstances the selected items (e.g., movies) are not intended for personal usage but rather for consumption in group. Group recommendations can assist a group of users in nding and selecting interesting items thereby considering the tastes of all group members. Traditionally, group recommendations are generated either by aggregating the group members' recommendations into a list of group recommendations or by aggregating the group members' preferences (as expressed by ratings) into a group model, which is then used to calculate group recommendations. This paper presents a new data aggregation strategy for generating group recommendations by combining the two existing aggregation strategies. The proposed aggregation strategy outperforms each individual strategy for di erent sizes of the group and in combination with various recommendation algorithms.
H.3.3 [Information Search and Retrieval]: Information Filtering; H.5.3 [Information Interfaces and Presentation]: Group and Organization Interfaces
Although the majority of the currently deployed
recommender systems are designed to generate personal
suggestions for individual users, in many cases content is selected
and consumed by groups of users rather than by
individuals. This strengthens the need for group recommendations,
providing suggestions thereby considering the tastes of all
group members. In the literature, group recommendations
have mostly been generated by one of the following two data
aggregation strategies [
The rst aggregation strategy (aggregating
recommendations) generates recommendations for each individual user
using a general recommendation algorithm. Subsequently,
the recommendation lists of all group members are
aggregated into a group recommendation list, which (hopefully)
satis es all group members. Di erent approaches to
aggregate the recommendation lists have been proposed during
the last decade, such as least misery and plurality voting [
The second grouping strategy (aggregating preferences)
combines the users' preferences into group preferences. This
way, the opinions and preferences of individual group
members constitute a group preference model re ecting the
interests of all members. Again, the members' preferences
can be aggregated in di erent ways, e.g., by calculating the
rating of the group as the average of the group members'
ratings [
This paper presents a new data aggregation strategy, which
combines the two existing strategies and outperforms each of
them in terms of accuracy. For both individual data
aggregation strategies, we used the average function to combine
the individual preferences or recommendations. Although a
switching scheme between both aggregation strategies has
already been investigated [
A major issue in the domain of group recommender systems is the evaluation of the accuracy, i.e., comparing the generated recommendations for a group with the true preferences of the group. Performing online evaluations or interviewing groups can be partial solutions but are not feasible on a large scale or to extensively test alternative con gurations. For example, in Section 5, ve recommendation algorithms in combination with two data aggregation strategies are evaluated for twelve di erent group sizes, thereby leading to 120 di erent setups of the experiment. Therefore, we are forced to perform an o ine evaluation, in which synthetic groups are sampled from the users of a traditional singleuser data set. Since movies are often watched in group, we used the MovieLens (100K) data set for this evaluation.
In the literature, group recommendations have been
evaluated several times by using a data set with simulated groups
of users. Baltrunas et al. [
This o ine evaluation is based on the traditional procedure of dividing the data set in two parts: the training set, which is used as input for the algorithm to generate the recommendations, and the test set, which is used to evaluate the recommendations. In this experiment, we ordered the ratings chronologically and assigned the oldest 60% to the training set and the most recent 40% to the test set, as this re ects a realistic scenario the best.
The used evaluation procedure was adopted from
Baltrunas et al. [
The e ectiveness of the di erent aggregation strategies
is measured for di erent sizes of the group and in
combination with various state-of-the art recommendation
algorithms. The used implementation of Collaborative
Filtering (CF) is based on the work of Breese et al [
Previous research [
Therefore, we propose an advanced data aggregation strategy which combines both individual aggregation strategies thereby yielding an accuracy gain compared to each individual aggregation strategy for di erent recommendation algorithms. This combination of strategies aggregates the preferences of the users as well as their recommendations with the aim of merging the knowledge of the two aggregation strategies into a nal group recommendation list. The idea is that if one of the aggregation strategies comes up with a less suitable or undesirable group recommendation, the other aggregation strategy can correct this mistake. This makes the group recommendations resulting from the combination of strategies more robust than the group recommendations based on a single aggregation strategy.
The two aggregation strategies are combined as follows. First, group recommendations are calculated by using the selected recommendation algorithm and the aggregating preferences strategy. The result is a list of all items, ordered according to their prediction score. In case of an individual aggregation strategy, the top-N items on that list are selected as suggestions for the group. After calculating the group recommendations using the aggregating preferences strategy, or in parallel with it, group recommendations are generated using the chosen algorithm and the aggregating recommendations strategy. Again, the result is an ordered list of items with their corresponding prediction score.
Both of these lists with group recommendation can still contain items that are less suitable for the group, even at the top of the list. The next phase will try to eliminate these items by comparing the two resulting recommendation lists. Items that are at the top of both lists are probably interesting recommendations, whereas items at the bottom of both lists are usually less suitable for the group. Less certainty exists about the items that are at the top of the recommendation list that is generated by one of the aggregation strategies but that are in the middle or even at the bottom of the recommendation list produced by using the other aggregation strategy. Therefore, both recommendation lists are adapted by eliminating these uncertain items in order to contain only items that appear at the top of both recommendation lists, thereby reducing the risk of recommending undesirable or less suitable items to the group. So, items that are ranked below a certain threshold position in (at least) one of the recommendation lists generated by the two aggregation strategies, are removed from both lists. If only one aggregation strategy is used, identifying uncertain items based on the results of a complementary recommendation list is not possible. In this experiment, we opted to exclude these items from the recommendation lists, that are not in the top-5% of both recommendation lists (i.e., the top-84 of recommended items for the MovieLens data set). As a result, the recommendation lists contains only items that are identi ed as `the most suitable' by both aggregation strategies, ordered according to the prediction scores calculated using either the aggregating preferences strategy or the aggregating recommendations strategy.
Subsequently, the two recommendation lists are combined into one recommendation list by combining the prediction scores of each aggregation strategy per item. In this experiment, we opted for the average as method to combine the prediction scores. So in the resulting recommendation list, each item's prediction score is the average of the item's prediction score generated by the aggregating preferences strategy and the item's prediction score produced by the aggregating recommendations strategy. Alternative combining methods are also possible, e.g., a weighted average of the prediction scores with weights depending on the performance of each individual aggregation strategy. Then, the items are ordered by their new prediction score in order to obtain the nal list of group recommendations. 5.
Our combined aggregation strategy is compared to the
individual aggregation strategies in Figure 1. Since users are
randomly combined into groups and the accuracy of group
recommendations is depending on the composition of the
groups, the accuracy slightly varies for each partitioning of
the users into groups. (Except for the partitioning of the
users into groups of 1 member, which is only possible in 1
way.) Therefore, the process of composing groups by taking
a random selection of users is repeated 30 times and just
as much measurements of the accuracy are performed. So,
the graph shows the mean accuracy of these measurements
as an estimation of the quality of the group
recommendations (on the vertical axis), as well as the 95% con dence
interval of the mean value, in relation to the
recommendation algorithm, aggregation strategy, and the group size.
The group size is indicated on the horizonal axis. The
vertical axis crosses the horizontal axis at the quality level of
the most-popular recommender. The pre x \Combined" of
the bar series stands for the proposed aggregation strategy
which combines the aggregating preferences and
aggregating recommendations strategy. The pre x \Pref" and \Rec"
indicate the accuracy of the two individual strategies,
respectively the aggregating preferences and aggregating
recommendations strategy. For each algorithm, only the most
accurate individual strategy is shown: aggregating
preferences for UBCF and CB, aggregating recommendations for
SVD, IBCF, and Hybrid [
The non-overlapping con dence intervals indicate a signi cant improvement of the combined aggregation strategy compared to the best individual aggregation strategy. Table 1 shows the results of the statistical T-tests comparing the mean accuracy of the recommendations generated by the best individual aggregation strategy and by the combined aggregation strategy for groups with size = 5. (Similar results are obtained for other group sizes.) The null hypothesis, H0 = the mean accuracy of the recommendations generated by using the best individual aggregation strategy is equal to the mean accuracy of the recommendations generated by using the combined aggregation strategy. The small p-values (all smaller than 0.05) prove the signi cant accuracy improvement of our proposed aggregation strategy.
This paper presents a new strategy to aggregate the tastes of multiple users in order to generate group recommenda0.905
0.9 0.895 G C nD0.89 n a e M 0.885 0.88 0.875 0.87
Comparison of the best individual aggregation strategy and the combined aggregation strategy tions. Both existing data aggregation strategies are combined to make a more informed decision hereby reducing the risk of recommending undesirable or less suitable items to the group. The results show that the combination of aggregation strategies outperforms the individual aggregation strategies for various sizes of the group and in combination with various recommendation algorithms. The proposed aggregation strategy can be used to increase the accuracy of (commercial) group recommender systems.