K-PropNet: Knowledge-Enhanced Hybrid Heterogeneous Homogeneous Propagation Network for Recommender System

In order to address the cold start problem and enhance model interpretability, recommender systems commonly incorporate knowledge graphs as supplementary information. However, existing knowledge graph-based recommendation methods primarily focus on exploring users’ deeper-level preferences or establishing item associations based on implicit knowledge graph information. Unfortunately, these approaches tend to overlook the crucial interactions between users and items, resulting in the underutilization of the rich information available from both parties. To overcome these challenges, this paper introduces a recommender model called the K-PropNet: Knowledge-Enhanced Hybrid Heterogeneous Homogeneous Propagation Network for Recommender System, which combines heterogeneous and homogeneous propagation techniques. Heterogeneous propagation leverages explicit user-item interaction information to broaden user interests, while homogeneous propagation employs convolutional neural networks to incorporate the structural and semantic information inherent to items themselves. Furthermore, the model utilizes two neural networks with attention-like mechanisms to discern the respective contributions of various neighbors in the two propagation paths within the knowledge graph. Ultimately, the information from both heterogeneous and homogeneous propagation is integrated to make the final prediction. The effectiveness of the proposed K-PropNet model is evaluated on four public datasets, demonstrating a significant performance improvement over state-of-the-art baseline methods.