A motif based hypergraph multi-level semantic encoding framework for social recommender systems

Hangyuan Du, Wenjian Wang, Liang Bai, Lu Bai, Jiye Liang

Abstract: Introducing social relations to improve recommendation quality, namely social recommender systems (SocialRS), has become a hot topic in academic and industrial communities. While many promising models have been proposed, most of them focus on mining latent user preferences from pairwise relations. However, in many real-world applications, user preferences are always implicit in both low-order relations and high-order interactive patterns. To address this issue, we develop a novel SocialRS framework, called Hypergraph Multi-level Semantic Encoding SocialRS (HMSE-SR). Specifically, the model first constructs multi-view hypergraphs conditioned on different types of motifs. On this basis, we propose a multi-level embedding learning paradigm that integrates the local interactive relation encoder with global hypergraph structure learning, so as to comprehensively mine latent user preferences from both low-level and high-level semantic levels. Furthermore, to overcome the problem caused by scarcity and skewed distribution of user relations in reality, we enhance the hypergraph encoder via distilling self-supervision signals across the local and global structure levels. Finally, a joint optimization model is developed to train the HMSE-SR. To verify the superiority of the proposed model, extensive experiments are conducted on four real-world datasets under both general and cold-start settings.

External IDs:dblp:journals/sigpro/DuWBBL25