Boosting Knowledge Graph with Diverse-Aware Intent Inference for recommendations

Knowledge graphs (KGs) have demonstrated significant effectiveness in recommendation systems due to their rich semantic structure. To overcome the limitations of traditional collaborative filtering and embedding-based methods, Graph Neural Network (GNN)-based approaches have been introduced to model the complex relationships within KGs. However, existing GNN-based methods face two key challenges: (1) they aggregate information indiscriminately from all neighboring nodes, leading to redundancy and inefficiency, and (2) they often prioritize similar items, which limits recommendation diversity and overall system performance. To address these challenges, we propose a Knowledge Graph with Diverse-Aware Intent Inference (KGDII), a novel framework that enhances both the quality and diversity of recommendations. KGDII generates user intents-representing users' underlying goals or preferences-by selecting a diverse subset of relationships within the KG. An attention mechanism assigns higher importance to more significant relationships, enabling the framework to produce more informative and diverse intent representations while reducing redundancy. Extensive experiments on real-world datasets show that KGDII outperforms state-of-the-art methods in recommendation accuracy and diversity. Ablation studies and case analyses further highlight the strong interpretability of KGDII, making it a promising approach for advancing recommendation system performance.