Preference Management in a Single User Group Recommender System

Groups often need to identify a product or service to consume together. For instance, when a group of friends is planning a vacation, they may look for an itinerary that satisfies all their wishes and wants. Group Recommender Systems (GRSs) employ special recommendation techniques designed to aid groups in their decision-making process. They can acquire the group members’ preferences and suggest items that satisfy them as much as possible [1].

In some decision-making scenarios, group members actively participate in the process, by expressing their preferences and discussing alternative options. However, sometimes just a single group member, here called the organizer, may be involved in evaluating options and making a choice, by considering all the group members’ preferences, according to his/her knowledge of them. This is a challenging application scenario for GRSs that, despite the absence of the group members’ active participation to the preference elicitation, and options discussion phases, is not necessarily simpler to address.

When an organizer searches for a convenient restaurant for his/her group, the following tasks should be supported: (i) recalling or anticipating group members’ preferences, depending on the available knowledge of them; (ii) reconciling possibly incompatible preferences of the group members; and (iii) identifying a restaurant that offers at least one dish suited for each group member’s preferences.

To support these tasks we have designed a novel version of MyFoodGRS¹ [2, 3], a GRS that enables an organizer to set a group, enter and revise the group members’ preferences, evaluate restaurants’ qualities, and make an informed choice for his/her group. The food preferences of a group member are specified by the organizer as a set of preferred food categories, and each dish, i.e., a meal included in the menu of a restaurant, belongs to one category. We consider ten food categories: Pasta, Red Meat, Fish, Rice, Burger, Pizza, Salad, Soup, White Meat, Chinese Noodle.

MyFoodGRS also helps the organizer to recall or anticipate a group member’s preferences (i.e., make reasonable guesses), if she/he does not have this knowledge, by using demographic and preferred cuisines information that may be more easily available [4]. MyFoodGRS uses this information in a machine learning model as predictive features of the group member’s food preferences. Moreover, group members’ preferences may be incompatible: for instance, it could be hard to find a place that serves Pizza (preference of a member) and also Fish (preference of another member). Hence, by utilizing an association rules component, MyFoodGRS assists the organizer in expanding the already entered food preferences of the group members and discovering new alternative preferences of the group members. In summary, we make the following contributions:

In the rest of the paper, a brief overview of the GRSs state-of-the-art is provided in Sect. 2, in Sect. 3 we describe the usage of MyFoodGRS in the target application scenario. Then, Sect. 4 describes the implementation of the preference management and recommendation functions. The designed user study and its results are described in Sects. 5 and 6. We conclude the paper by discussing its limitations and future works in Sect. 7.

Recommender Systems (RSs) help their users to find content that meets their preferences and needs [5]. Group Recommender Systems (GRSs) are RSs that provide personalized content to a group of people; their main challenge is to properly aggregate the preferences of the group members, such that the recommendations satisfy everybody in the group [1].

A variety of GRSs have been proposed to address the specific requirements of the group recommendation task [1, 6]. However, we have found that only two, namely, Intrigue [7] and Rempad [8], are addressing an application scenario similar to that of MyFoodGRS, as they both support an organizer to make a decision on behalf of an entire group. Intrigue enables a tour leader to insert group members’ information and generate recommendations to satisfy different subgroups. Rempad supports organizers in selecting multimedia content for reminiscence therapy sessions. In comparison to Intrigue, which is suggesting a set of points of interest (POIs), recommending a single restaurant satisfying a wide spectrum of food preferences is harder, as more constraints must be satisfied. Moreover, differently from Rempad, in our scenario the organizer has a more limited knowledge of the individual group members’ preferences, hence requires some system support.

From another point of view, commercial restaurant finders, such as TripAdvisor, are designed for individuals seeking restaurants for their group, but they do not explicitly provide support functionalities such that the selected item actually fits all the group members’ needs and wants. In fact, none of them support the organizer in facing the two key tasks mentioned in the Introduction, namely, recalling and anticipating other group members’ preferences and helping the organizer in dealing with incompatible preferences.

Our approach is motivated by Logue and Smith [4], who introduce age, thinness, sensation seeking, and ethnic background (i.e., the main cuisine of the users on which they were raised) as predictors of food preferences for adult humans. They show that by using these features one can predict food preferences of individuals. They also model patterns of food preferences by establishing relations between them. For instance, they have found that the tendency to prefer junk dishes is correlated with a tendency to prefer meat and potatoes, and breakfast dishes. We have implemented these ideas in machine learning models to support the organizer to expand the initially entered group members’ preferences.

Let’s assume that Jack is arranging a dinner for a group including also Jim and Frank. The interaction with MyFoodGRS starts in the GROUP tab, where Jack can customize the system behavior and set up the group (Fig. 1a). Then, he can enter his and the other group members’ (known) food preferences, i.e., favorite food categories (Fig. 1b). The preferences of a group member are considered to be “alternative” ones, i.e., they are expressing OR conditions. The system assigns a color to each group member (e.g., red to Frank), that always indicates the group member and his/her preferences in the various system screens.

Jack, the organizer, can also access two preference management functionalities: one helps Jack to recall Jim or Frank’s preferences and another helps him to overcome the difficulty of finding a suitable restaurant in case of incompatible preferences. Recalling preferences highlights possible preferences of a group member as soon as he/she is introduced in the system by the organizer, whereas the incompatible preference suggestion support requires that an initial set of group members’ preferences have been entered by the organizer. To recall a group member’s preferences, Jack can click on “Suggestion based on member’s info” and input member information: nationality, gender, age, or preferred cuisines (Fig. 1e). Suggested preferences for a group member are indicated by adding a border, of the member’s distinguishing color, around preferences not yet attributed (hence in green) to the group member (Fig. 1f). Suggestions for overcoming incompatible preferences are also indicated in a visually similar way.

The organizer (Jack), after having entered some group members’ preferences, can access the system’s recommendations in the RESTAURANT tab (Fig. 1c), where restaurants are ranked based on their “Fitness to the group” (see Sect. 4 for the technical details). For a better assessment of the restaurant’s suitability for the group, clicking on its card (Fig. 1c) displays restaurant details: (a) popularity of the restaurant in the system; (b) fitness for the group, showing the predicted group score of the restaurant (see Sect. 4) and detailed information about the suitability of the restaurant for each group member; (c) similarity to the groups who bookmarked this restaurant in the past, that is, showing to what extent other groups who bookmarked this restaurant are similar to the organizer’s group; (d) TripAdvisor score of the restaurant; and (e) food categories and dishes diversity on the restaurant menu.

When the organizer reaches the last recommendation in the list that satisfies all the group members, if he is not satisfied with any of the recommended restaurants, is invited to use the two preference management supporting functionalities we mentioned above. Hence, by clicking on the REVISE button, the organizer is redirected to the GROUP tab, where the two preference management functionalities are available.

Finally, the organizer can bookmark some restaurants and access them by using the Bookmark tab. Bookmarked restaurants can be compared side-by-side and one of them can be finally chosen for the group.

To sort the recommendations, the organizer can select one of the three available ranking algorithms, while two machine learning models are supporting the organizer in, recalling and anticipating the other members’ preferences, and dealing with incompatible preferences.

We now focus on the evaluation of the usefulness of the two proposed supporting functionalities: helping the organizer to (i) recall and anticipate the preferences of the group members, and (ii) deal with incompatible preferences. To this end, we have conducted a user study, which will be now described, providing details on the design, i.e., independent and dependent variables, data collection procedure, and resulting measures. At the end, we will elaborate on the conducted analyses.

The user study was designed to assess whether the two proposed supporting functionalities enhance the organizers’ ability to make better decisions. The “goodness” of a decision is quantified with two constructs, (a) the individual loss, i.e., how well the group choice aligns with the group members’ individual preferences [10], and (b) the group score, i.e., how highly the group members rate the selected option as appropriate for the group. The operationalization of these constructs will be in detail described below. These constructs are the dependent variables of our analysis - we measure how good the choice of the organizer is, depending on the supporting functionalities in the system.

The realization of the independent variables is done with four variants of the system. Namely, \(V_N\) does not offer any of the two supporting functionalities. \(V_R\) provides support for Recalling preferences, \(V_I\) provides support for Overcoming Incompatible preferences, and finally, the variant \(V_{RI}\) provides both supporting functionalities. As an extraneous condition, that can affect the quality of the organizer’s choice, we have considered the level of information about the group members available to the organizer. To this end, we have defined three types of groups: (i) \(G_N\) - Groups with no information, i.e., only group members’ demographic information (age, gender, and nationality) is available, (ii) \(G_P\) - Groups with partial information, i.e., demographic information and preferred cuisines of the group members are available, and (iii) \(G_F\) - Groups with full information, i.e., demographic information, preferred cuisines, and preferred food categories are revealed to the organizer. During the experiment stage, we followed the A/B testing experimental design [11]. We note that for each group, all the group members played the role of “organizer”, they were provided with one of the introduced system variants, and they were shown different levels of information about their fellow group members.

Data Collection Procedure was organized in three stages, Before Decision-Making, Decision-Making, and After Decision-Making. In the Before Decision-Making stage, each participant was introduced to the overall procedure with an explanatory movie, and entered personal demographic information (age, gender, and nationality), preferred food categories, and preferred cuisines. The system also randomly selected for each user 10 restaurants such that each restaurant had at least one of the user’s food preferences in their menu, and asked the users to rate them based on their preferences.

After all the involved users had performed the Before Decision-Making stage, the system automatically constructed groups of different sizes (containing from 2 to 5 members). Then, started the Decision-Making stage where each group member was asked to independently interact with a variant of the system and to play the role of the organizer of the group. Note that a user was often assigned to more than one group. In this stage, the system also showed different levels of information about other group members to the organizer, depending on the group type to which a particular user was assigned. The task for the organizer was: Imagine that you are responsible for finding a proper restaurant for a group of people. This is a group of people that you might or might not know. Based on the information about the group members provided in the system, and with the help of the system’s functionalities, your task is to select one restaurant that you believe will make the group (all group members) happy. In parallel, all the other group members performed the same procedure. This stage was completed when all the organizers selected a restaurant for their group.

The participants were faculty members of the Free University of Bozen-Bolzano and bachelor students at the University of Sarajevo. Altogether, 95 participants were involved, and 60 groups were generated: each user participated in two groups and very few in three groups. However, not all the participants completed all the initiated evaluation sessions (195) in their groups, and we obtained 153 evaluations (26 users completed one evaluation, 53 two evaluations, and 7 three evaluations). In each evaluation, a user rated the choices made by the other group members when they acted as organizers, and overall we collected 492 ratings and group scores for restaurants chosen by group members acting as organizers. Participants were from 11 countries, mainly Bosnia, Italy, and Iran. 54% are males and 46% females, with an average age of 21. The minimum number of organizer’s choice ratings/group-scores that the combination, i.e., the variant of the system and group type (altogether 12 combinations), received was 25, the maximum was 56, with the mean of 41.

Analyses. In order to evaluate the effectiveness of the two proposed supporting functionalities in terms of the organizer’s choice quality, the Cumulative Link Model (CLM) [12] (also known as two-way ordinal regression, or proportional odds model [13]) with the help of R package ordinal [14, 15] was used. This analysis is the most similar to the factorial (two-way) ANOVA [16, 17], which is suitable when there is more than one independent, categorical variable, and the goal is to measure the effect of each variable as well as their interaction on a continuous dependent variable. However, in our case, the dependent variables are ordinal (not continuous), i.e., individual loss takes integer values from 0 to 4 (lower is better), and group score takes also integer values from 1 to 5 (higher is better), and for both, order of values is relevant. To this end, factorial ANOVA was not a suitable choice, and the CLM was instead utilized. The significant results are interpreted as “there is a significant effect of the independent variable on the dependent variable”. The assumption that CLM needs to meet in order to be applicable for the data is the proportional odds assumption [18], which in our case was not violated.

The output of the CLM is shown in Table 1, and indicates that the system variant has a significant effect on individual loss as well as on group score: there are differences in the quality of the organizer’s choice with respect to different variants of the system. Conversely, no significant effects were found for the group type (i.e., the level of information about the group members that was provided to the organizer), nor the interaction effect of the two independent variables. Therefore, we hypothesize that depending on the group type, different system variants provided different levels of support, however leading to the same quality levels of the organiser’s choice. Nevertheless, both models are significant, indicating that their predictive power is significantly better than that of models using only intercept values.

Next, since we have four variants of the system, we were interested in which variants actually excel in terms of the two dependent variables. Since the level of available information did not play a role in the quality of the organizer’s choice, we proceeded with a simpler analysis of variance for an ordinal (rank) variable, i.e., the Kruskal-Wallis test and the post-hoc Dunn test.

Kruskal-Wallis test indicates significant differences between the four system variants, for individual loss with p-value of 0.000, as well as for the group score variable with p-value of 0.019 (which was expected given the previously presented results). The results of the Dunn test are provided in Table 2 (the reported p-values are adjusted for multiple comparisons using the Bonferoni correction [19]. The table only shows significant differences due to the page limit. It is found that individual loss is lower and group score is higher when the organizer’s choice was made within the variant that provides support for both recalling/anticipating and dealing with incompatible preferences, in comparison to the organizer’s choices made within the variant with no supporting functionalities, and the variant which only provided support for recalling group members’ preferences. Furthermore, the individual loss was significantly lower for the organizer’s choice when the support for incompatible preferences was provided in comparison to when support for recalling preferences was given. Here, no significant differences were observed for the group score. These differences are also visualized in Fig. 2.

We have here presented MyFoodGRS; it assists a group member (organizer) to choose a restaurant on behalf of the entire group. We have evaluated two supporting functionalities: recalling and anticipating group members’ preferences, and dealing with possibly incompatible preferences within the group. The first one utilizes a machine learning model that predicts the most probable member’s individual preferences based on available demographics or preferred cuisine data. The second one makes use of association rule mining to extend the preferences of a group member entered by the organizer. In our ablation analysis, we found that the variant offering both supporting functionalities resulted in significantly better choices, i.e., choices for which individual loss of group members is significantly lower, and their rating for the choice is significantly higher, in comparison to choices made with variants of the system that do not provide these functionalities. Moreover, the variant that only provides support for recalling and anticipating group members’ preferences did not result in significantly better choices.

We would like to highlight some of the limitations of this study. First of all, we acknowledge that the number of food categories that we used (only 10) is a limiting factor in fully expressing food related preferences. Additionally, our participants were mostly bachelor students and may not be representative of real-world users. Furthermore, we expected to observe certain differences in the usefulness of the proposed supporting functionalities in relation to the level of available information about the group members, but this was not the case, and in order to better understand this, additional analysis, with an extended data sample is necessary. We speculate, that the recalling functionality may have not been effective for at least three reasons. First, because of a sub-optimal machine learning model that we have used and the limited training data. Secondly, because of the low variability of the users’ demographics in our data sample, as most of the participants are from Bosnia and of similar age. Third, the usefulness of the proposed functionality may have been reduced by the small number of food categories.

We acknowledge as well that randomly formed groups do not represent typical groups, which tend to be composed by more similar individuals. As part of our future work, we aim to conduct a larger experiment with established groups. This new experiment is also intended to be a between-subjects study, which requires more data samples but is considered as more reliable.