Abstract: Recommendation systems are a crucial means to address the problem of information overload, and knowledge graphs serve as the most widely used auxiliary information source in recommendation systems. Existing knowledge graph-based approaches suffer from limitations including insufficient dynamic modeling of multi-semantic relationships, lack of modeling for neighbor sequence dependencies and long-range associations, and inadequate capability in extracting multi-level features from graph structures. To tackle these issues, this paper proposes a Multi-Semantic and Temporal Convolutional model for Knowledge Graph recommendation (MSTC-KG). The model dynamically fuses user-relation interactions through a multi-head attention mechanism and a gating mechanism to distinguish the semantic contributions of neighbor nodes, employs GRU to capture the temporal dependencies and long-range associations of neighbor sequences, and utilizes 1D convolutional neural networks to extract local-to-global multi-level features of graph structures. Compared with 14 baseline models on two public datasets, MSTC-KG achieves an AUC of 0.737 and 0.982, and an F₁-score of 0.680 and 0.938 on the Book-Crossing and MovieLens-20M datasets respectively, outperforming all other baseline models. Experiments demonstrate that through the architecture characterized by “dynamic multi-semantic differentiation + temporal dependency capture + hierarchical feature extraction”, the model effectively enhances recommendation accuracy and the capability to model complex graph structures.

Key words: knowledge graph, recommendation system, multi-head attention, GRU, convolutional neural network