Knowledge-aware fine-grained attention networks with refined knowledge graph embedding for personalized recommendation

Recommendation systems aim to provide users with personalized and accurate services by integrating various machine learning technologies. Suffering from the puzzles such as cold-start and data sparsity, recommendation models has received extensive attention by incorporating knowledge graphs as additional supplementary information to effectively solve these problems. More recently, graph neural networks (GNNs) have been adopted to establish knowledge-aware recommendation models and made considerable achievements. Nevertheless, the existing GNN-based approaches are inadequate in the following two aspects: (1) How to achieve sufficient high-order collaborative signals? (2) How to reduce the impact of redundant KG information on representation? To overcome these limitations, we propose a novel framework KFGAN with Knowledge-aware Fine-grained Attention Networks for personalized knowledge-aware recommendation, which captures user's preferences by encoding the relation paths and associated entities and generates refined knowledge graphs to learn the potential semantic information. Specifically, by integrating the high-order collaborative signals of users and items and the structural information of knowledge graph, KFGAN enriches the feature representation of users and items and realizes the consistency and coherence of CF and KG information. Furthermore, KFGAN draws a lesson from graph contrastive learning approach to accomplish refined knowledge graph embedding, which alleviates the interference of redundant KG signal to the model and mines the latent semantic information in KG. Massive experimental results on three benchmark datasets prove that KFGAN model dramatically outperforms the current state-of-the-art baselines. The code and experimental datasets will be available at https://github.com/weiwang1992/KFGAN to verify and study for further research.