KHG-Aclair: Knowledge Hypergraph-Based Attention with Contrastive Learning for Recommendations

Information overload and complex user interactions make it difficult to retrieve valuable data. Recommendation systems have become crucial in addressing this challenge by providing users with relevant information and items. Collaborative filtering-based recommendation methods, which are commonly used in this context, often suffer from data scarcity, thus limiting their effectiveness for users with insufficient interaction data. To overcome this problem, knowledge graphs have been integrated into recommendation systems to enhance user and item representation through the implementation of additional semantic relatedness. Despite their potential utility, most recommendation models assume binary relations within knowledge graphs, thereby overlooking the high-order relationships that are prevalent in knowledge graphs. Knowledge hypergraphs, which can capture complex and multi-dimensional relationships, offer a solution to this limitation. This paper proposes KHG-Aclair, a novel recommendation system that leverages hypergraphs to uncover hidden features within knowledge graphs, thus enhancing recommendation accuracy and insight. We have transformed the Freebase knowledge graph into a knowledge hypergraph and made this dataset publicly available. KHG-Aclair also incorporates contrastive learning to refine the knowledge hypergraph, thus reducing noise and improving representation for less popular items. Altogether, our model demonstrates strong generalizability, as it achieves high performance across multiple datasets, thus indicating that it can serve as a versatile solution for various recommendation systems. Our implementation codes are available at https://github.com/HBD-NGC1316/KHG-Aclair.