Beyond accuracy-centric ranking: A collaborative knowledge graph–guided deep learning recommendation framework balancing accuracy and diversity

Personalized recommendation has become an indispensable installation through which e-commerce platforms improve matching efficiency and enhance user experience. However, recommendation models that focus primarily on ranking accuracy may also narrow product exposure and generate repetitive recommendation lists. Although knowledge-graph-enhanced recommendation provides a useful way to enrich semantic information and alleviate sparsity, existing methods still face challenges in balancing recommendation relevance and diversity, especially under sparse and long-tail conditions. We propose KMAGNN, a collaborative knowledge graph-based recommendation framework that integrates relation-aware multi-head attention, gated residual propagation, and diversity-aware neighborhood regulation. The proposed framework captures heterogeneous semantic relations among users, items, and entities while stabilizing high-order message passing and reducing redundant semantic aggregation. In addition, dynamic negative sampling and a joint optimization scheme are introduced to support a more balanced recommendation objective. Experiments on multiple public datasets show that KMAGNN achieves a more favorable accuracy–diversity trade-off than competing methods and remains robust under sparse interaction settings. This study extends knowledge-graph-based recommendation by demonstrating how structured semantic modeling can support recommendation outcomes that are both relevant and diverse.