Heterogeneous Propagation-Based Recommendation Systems with Knowledge Graph Attention Network

Collaborative filtering-based recommendation systems often face challenges such as data sparsity and cold start problem. Knowledge graphs have been leveraged as auxiliary information to mitigate these issues. However, current knowledge graph-based recommendation systems still struggle to accurately capture the latent features of users and items in the embedding space. In this article, we introduce a heterogeneous propagation-based recommendation framework that effectively integrates collaborative patterns derived from user–item interactions with knowledge graph data. Specifically, we first design hierarchical layerwise attention mechanisms (LAMs) tailored to the different layers of the knowledge graph propagation process. These mechanisms assign appropriate attention scores to feature representations at various levels of complexity, thereby enhancing the integration. Additionally, the feature learning is further improved by adding Gaussian noise to the initial latent features of users and items, which are then refined using a denoising autoencoder. This method helps learn more robust and discriminative representations for users and items, enhancing recommendation performance. Extensive experiments were conducted on the MovieLens-1M, Book-Crossing, and Last.FM datasets demonstrate that the proposed algorithm significantly outperforms eighteen state-of-the-art baselines. Furthermore, ablation studies confirm the effectiveness of each module, emphasizing their individual contributions to the overall performance improvement.