Abstract: Knowledge graphs have been widely applied in recommendation systems, contributing to alleviating the data sparsity issue in user-item interactions. Existing knowledge graph-based recommendation methods primarily rely on path mining or information propagation to explore potential associations between users and items. However, they often fail to fully leverage the rich semantics and structural information within knowledge graphs, and may introduce irrelevant noise, thereby affecting the accuracy of recommendations. To address this, we propose a recommendation method that integrates hierarchical knowledge graph embedding with self-attention mechanism, aiming to extract high-order semantics and structural information from knowledge graphs to mitigate the sparsity problem. Specifically, the proposed model first constructs high-order subgraphs for user-item pairs, explicitly depicting the complex relationships between them. Through a hierarchical attention embedding learning process, the model encodes the high-order semantics and topological structure within the subgraphs, and employs a self-attention mechanism to differentiate the importance of each entity in the subgraphs, ultimately generating high-quality subgraph embeddings for accurate user preference modeling. Experimental results on three real-world datasets show that the proposed method achieves an average improvement of 10.7% and 13.6% in Hit and NDCG metrics, respectively, compared to the state-of-the-art baseline models. Moreover, the proposed method consistently outperforms in scenarios with varying degrees of data sparsity, effectively alleviating the data sparsity issue.

Key words: Recommendation Systems, Graph Neural Networks, Knowledge Graphs, Self-Attention Mechanism, Collaborative Filtering

摘要： 知识图谱在推荐系统中已广泛应用，有助于缓解用户与项目交互数据稀疏性问题。现有基于知识图谱的推荐方法主要依赖路径挖掘或信息传播来探索用户与项目之间的潜在关联，但未能充分利用知识图谱中的丰富语义和结构信息，且易引入无关噪声，影响推荐的准确性。为此，提出了一种融合层级知识图谱嵌入与注意力机制的推荐方法，旨在挖掘知识图谱的高阶语义和结构信息以缓解数据稀疏性问题。具体而言，首先构建用户与项目的高阶子图，显式刻画用户与项目之间的复杂关联关系。通过层级子图嵌入模块，对子图中的高阶语义和拓扑结构进行编码，并结合自注意力机制，区分子图中各实体的重要性，最终生成高质量的子图嵌入表示，实现对用户偏好的精准建模。实验结果表明，在三个真实数据集上，相较于现有最优基线模型，所提方法在Hit和NDCG指标上分别实现了平均10.7%和13.6%的提升，且在不同数据稀疏程度的场景下推荐效果均占优，有效缓解了数据稀疏性问题。

关键词: 推荐系统, 图神经网络, 知识图谱, 自注意力机制, 协同过滤