Group recommendation Recommender systems by borattoludovico

Integrating Collaboration and Leadership in Conversational Group Recommender Systems

Interaction between the users, in a group setting, can support the decision-making when group recommendations have to be produced. Specifically, the presence of collaborative and leader users leads the group to trust these users when a final decision has to be taken.

In a paper published in the ACM Transactions on Information Systems (TOIS), with David Contreras and Maria Salamó, we propose a collaborative model based on the social interactions that take place in a web-based conversational group recommender system. The collaborative model allows the group recommender to implicitly infer the different roles within the group, namely, collaborative and leader user(s). Moreover, it serves as the basis of several novel collaboration-based consensus strategies that integrate both individual and social interactions in the group recommendation process. One year after the publication of this paper, we were also invited to present this study at the ACM SIGIR ’22 conference.

Modeling interaction

For each user, we build an Individual user model (IM), based on the interaction between this user and the products. In order to account for the interaction between a user and the rest of the group, we also build a Collaborative user model (CM), which measures (i) the number of product suggestions to other group members, (ii) the number of viewed products by other group members, and (iii) the number of products stored as preferred by other group members. We combine these three quantities with a weighted average, to shape a score that tells us show collaborative is this user. The leader of a group is the user with the highest collaborative score.

Collaboration-based consensus strategies

We shape a series of consensus strategies, either considering the individual model or accounting for collaboration. Specifically, we shaped six strategies that are based on classic consensus strategies, either with or without accounting for collaboration.

  • Individual/Collaborative mean: average of the individual preferences, possibly weighted by the collaboration of each user;
  • Individual/Collaborative completeness: favors high scores while penalizing big differences between members, possibly weighted by the collaboration of each user;
  • Individual/Collaborative multiplicative: multiplication of the individual preferences, possibly weighted by the collaboration of each user.

Moreover, we propose two new strategies, based on our notion of leadership:

  • Maximize leader satisfaction: out of all the preferences of the group, choose the item preferred by the leader;
  • Leader selection: only choose the item preferred by the leader.

Live-user evaluation

We validated our proposal with a live-user evaluation, involving 68 participants, divided into 17 groups of 4 people. The products were in the skiing domain, which is widely studied in the group recommendation domain. Specifically, recommendations had to be produced for a group, by choosing among 153 European skiing locations; each location was characterized by 41 features. Users were allowed to express their preferences for the locations and their features by using the gCOACH (COllaborative Advisory CHannel for group recommendation) platform.

The results are very encouraging, from multiple perspectives. As the following table show, the user who was predicted as the leader by our score is also recognized as the leader by the users in the evaluation in the vast majority of the cases.

Moreover, we can summarize the following outcomes (readers can refer to our paper for a detailed analysis):

  • When analyzing the personality of the leaders, in terms of conflict style (namely, competing, cooperating, compromising, avoiding, and accommodating), there is not a predominant mode of conflict style in the leader analysis. However, it can be observed that the competing, cooperating, and accommodating styles are the most common ones for the leaders, covering 67% of the leader styles;
  • Our analysis of the consensus-based strategies was based on the fact that each user rated the recommended product for each consensus strategy. Therefore, we analyzed the average score obtained by the items recommended following each of the strategies. The proposed strategies that use both the individual and collaborative models obtained better results than the ones that only use individual models. In addition, the strategies that maximize the interest of the leader reached the best results.