UDA: A user-difference attention for group recommendation

Human beings are gregarious by nature, and thus, group activities are indispensable in people's daily lives. In light of this, group recommendation systems have attracted wide research attention in recent years. The pivotal task of group recommendation is to learn the weights of group members and then aggregate their preferences. Most existing methods only consider a single user and the target item to calculate their weight, which is insufficient to determine the user's importance in a group. In this study, we emphasize the importance of comparison information. A comparison is an explicit relationship between users. Motivated by our observation and based on attentive group recommendation, we propose a user-difference attention (UDA) model that explicitly simulates the comparisons between group members using relational attention, which is different from previous single user-based models. Each user is compared with all other users under the guidance of the target item, and a multi-layer perceptron is then exploited to add nonlinear transformations. We propose several user relational kernels (URKs) to simulate different types of relations during group decision-making. Extensive experiments were conducted on three public datasets. The results show that UDA significantly exceeds the state-of-the-art competing methods. (c) 2021 Elsevier Inc. All rights reserved.

Automated summary

Group recommendation systems have gained significant research attention in recent years, with the focus on learning the weights of group members and aggregating their preferences. This study emphasizes the importance of comparison information and proposes a user-difference attention model to explicitly simulate comparisons between group members.