A cold-start aware group recommendation method based on an augmented graph and knowledge distillation

초록

Capturing a group consensus for group recommendations is a highly challenging problem. In the real world, many groups are often ad-hoc, so recommender systems frequently face a cold-start problem. Recent multi-view-based approaches, however, overlook unseen groups that have no historical data on items. Their performance could degrade when there are many unseen groups, because unseen groups do not have item-level information. Though there have been several cold-start-aware methods, they do not fully utilize higher-order relationships, e.g., group memberships. Hypergraph-based approaches have been shown to be effective for modeling higher-order relationships. However, there is no recursive hypergraph-based approach to cope with unseen groups yet. We develop a unified graph for group recommendations through a novel recursivehypergraph modeling that utilizes connected edges, e.g., group memberships and user preferences, so that we can obtain rich information for unseen groups. Since training on a unified graph alone is insufficient for generalization, we propose two regularization components based on knowledge distillation with graph augmentation for informative training of group preferences. One is to balance the information between unified and augmented graphs. The other is to distill members' preferences into groups' preferences. We conduct performance experiments using four real-world datasets and show that our proposed method significantly outperforms other recent group recommendation techniques on average.

키워드