Group Recommendation via Self-Attention and Collaborative Metric Learning Model

Group recommendation has attracted wide attention owing to its significance in real applications. One of the big challenges for group recommendation systems is how to integrate individual preferences of each group member and attain overall preferences for the group. Most of the traditional group recommendation solutions regard group members as equal participants and assign a same weight to each member. As a result, performance of this type of recommendation methods is not as good as expected. To improve the performance of group recommendation, a novel group recommendation model via Self-Attention and Collaborative Metric Learning (SACML) is presented in this paper. With the employment of Self-Attention mechanism, the SACML model can learn the similarity interactions between group members and services and decide a different weight for different group member. Based on these weights, group preferences for services can be generated by the aggregation of group members' preferences and the group's own preference. Similar metric space between group and services is obtained via collaborative metric learning with the group preferences and positive and negative services' features. Group recommendation is finally implemented based on the obtained metric space. Simulation has been conducted on CAMRa2011 and Meetup datasets, and experimental results show that the proposed SACML model has better performance in comparison with those baseline methods.