Structure-aware knowledge-graph attention for explainable recommendation under sparse user interactions

Abstract

The proliferation of digital content has amplified the reliance on recommender systems to filter information and deliver personalized user experiences. However, traditional collaborative filtering approaches suffer significantly from data sparsity and cold-start problems, particularly in environments where user interactions are highly limited. Furthermore, the opaque nature of complex neural recommendation models poses challenges for user trust and system transparency. To address these interconnected challenges, this paper proposes a novel framework centered on structure-aware knowledge-graph attention for explainable recommendation under sparse user interactions. By integrating auxiliary knowledge graphs, the proposed methodology maps items to a rich network of semantic relations, thereby supplementing the sparse user-item interaction matrix with dense factual connectivity. We introduce a sophisticated attention mechanism that is explicitly aware of the local and global topological structure of the knowledge graph. This mechanism adaptively weights different relational paths between users and items based on network motifs and path connectivity degrees, ensuring that the propagation of embeddings prioritizes the most informative and contextually relevant semantic links. Our framework not only improves recommendation accuracy in highly sparse scenarios but also provides explicit path-based explanations for every recommendation generated. Extensive theoretical analysis and structural evaluations demonstrate that the model successfully mitigates the sparsity issue while generating high-quality, intuitive explanations for end-users.