Multi-level Noise Filtering and Preference Propagation Enhanced Knowledge Graph Recommendation.

Knowledge Graph (KG) can provide semantic information about items, which can be used to mitigate the sparsity problem in recommendation systems. In recent years, the trend in knowledge-aware recommendation methods has been to leverage Graph Neural Networks (GNNs) to aggregate node information in KG. However, many of these methods focus on mining the item knowledge association on KG, but ignore the potential item auxiliary information in user’s history interaction outside KG. Furthermore, these methods equally aggregate all neighbor entities of the item on KG, which will inevitably introduce irrelevant entity-interaction behaviors. To address these issues, we propose a novel model, called Multi-level Noise Filtering and Preference Propagation Enhanced Recommendation (MNFP). Technically, we employ self-attention mechanisms to model the user’s interaction sequence to mine the item’s auxiliary information. Then, we design a twin-tower preference propagation mechanism that iteratively expands item auxiliary information on KG. Additionally, we propose a multi-level noise filtering mechanism. By learning the relationship consistency between the item and its neighbor entities, the model can guide the item to selectively link highly related neighbors in preference propagation, thus reducing the introduction of noise. We evaluate MNFP on three real-world datasets: MovieLens-1M, Last.FM and Book-Crossing. Results show that MNFP significantly outperforms state-of-the-art methods on AUC and F1 .