Decoupled Knowledge Propagation Network and Sampling Strategies for Recommendation System

State-of-the-art knowledge-graph-based models, such as RippleNet and CKAN, have been successfully applied to recommendation systems. These models make use of knowledge graph to expand the entity information, which is similar to ripples, so as to strengthen the correlation between the user's preference and the candidate item. In RippleNet and CKAN, however, the networks for learning the features of user preference and candidate item are mutually coupled, which might limit the recommendation performance, especially for multi-levels propagation. On the other hand, when entities are expanded over the knowledge graph, the amount of entities increases exponentially with the number of diffusion levels. In earlier research, random sampling was adopted to restrict the number of triples, which however may limit the coverage of the expanded knowledge for model learning as well as the prediction performance. To tackle these problems, we propose a decoupled knowledge propagation network that separates the user's preference and the candidate item to distinguish the candidate items better. Experiments conducted on the MovieLens 1M recommendation dataset and the Microsoft Satori knowledge graph show the proposed mode outperforms RippleNet and CKAN. In addition, three sampling strategies, including balanced sampling, non-duplication, and dynamic sampling were proposed to deal with the sampling issue, and experiments show such sampling strategies are effective and superimposed for the prediction performance.