MEDFA: A Knowledge-Aware Recommendation Model With Multi-Space Embedding, Feature Distillation, and Fused Attention

Abstract

Knowledge graphs have provided a new research direction for recommender systems. Although contrastive learning and knowledge graph embedding have achieved significant success in knowledge-aware recommendation, two key problems persist. First, while introducing contrastive learning enhances performance, it also incurs substantial model training costs. Second, the coexistence of both simple and complex data structures within recommendation system Knowledge graphs makes it difficult to model them effectively, either within a single geometric space or through multi-space approaches that simply sum the recommendation scores. To address these problems, this paper proposes a teacher-student framework incorporating Multi-space Embedding, Feature Distillation, and Fused Attention. Our method trains a student model that, through knowledge distillation from a pre-trained teacher, specifically learns the teacher’s embedding representations at different semantic levels, allowing it to inherit representational capabilities while reducing complexity. Furthermore, our method employs a multi-space attention fusion mechanism that adaptively leverages the distinct advantages of Euclidean, hyperbolic, and complex spaces to effectively model diverse data structures. We conducted extensive experiments on three public datasets. The results demonstrate that the student model not only outperforms all baseline models but also achieves the teacher model’s performance while being more efficient, reducing training time by 85%. Our experiments also provide an empirical justification for our model, demonstrating that Euclidean space is better suited for simple data structures, while hyperbolic and complex spaces excel with complex ones.