Recommendation Algorithm for Multi-View Contrastive Learning Based on Knowledge Graph

Abstract

Most of the existing knowledge graph-based recommendations only consider introducing information from the item side to improve the recommendation effect, ignoring the importance of social relationships in the recommendation process, and the deeper intentions behind the user-item interactions are not sufficiently considered when utilizing knowledge graphs for recommendation, which may result in the model learning suboptimal results, and most of the knowledge-based recommendations are under supervised learning paradigm, which has the problem of sparsely-supervised signals. Therefore, we propose a multi-view comparative learning knowledge graph recommendation method-MIKGR, which discovers user-item relationships from the granularity of intentions on the global structure, couples them with KG relationships, and introduces social relationships into the recommendation model to further alleviate recommendation data sparsity. The experiments are validated on three real datasets, showing that our model MIKGR is effective.