Knowledge-aware neighbor collaborative multi-relationship multi-interest comparative recommender system for diversified book recommendations

Book recommendation system has become an essential tool for library information transformation, and has received widespread attention recently. Emerging graph neural networks and knowledge graphs are used in today’s recommendation systems for their benefits to data mining. Although some related progress has been made, many existing researches mainly focus on recommendation accuracy too much and miss diversity. In addition, the researches that primarily focus on the interaction between users and items, overlook the multi-attribute information between users and items, which limits the collaborative signals between entity attributes, resulting in unsatisfactory recommendation effects. To address these shortcomings, an Entity-attribute Interactive Knowledge Graph (EIKG) is constructed, in which readers, books, and rich side attributes are unified through typed edges, allowing attribute semantics to propagate alongside behavioral links. This paper aims to embed attribute collaborative signals into the recommendation tasks by utilizing multi-attribute knowledge graphs of readers and books, thereby improving the accuracy and diversity of recommendation systems. Building on the EIKG, a knowledge-aware book recommendation system is proposed. Specifically, four critical parts are designed for the model: 1) a multi-relation attention component that automatically highlights semantically important edges and suppresses noisy ones, 2) the neighbor attribute collaborative component that improves reader behavior similarity and book theme consistency, 3) an attribute-guided contrastive objective that explicitly pulls together diverse themes while retaining highly relevant titles, 4) the reader multi-interest channel adaptively generates multiple interest embeddings based on the reader’s borrowing history data to tailor recommended books for each reader. The seamless coupling of these components forms a unified end-to-end framework that jointly optimizes accuracy and topically aware diversity. By applying these four components, an emerging multi-task training framework that considers recommendation accuracy and book theme diversity is constructed. Experiments on practical datasets show that the model has high recommendation recall and topic diversity.