Knowledge-Enhanced Top-K Recommendation in Poincaré Ball

Personalized recommender systems are increasingly important as more content and services become available users struggle to identify what might interest them. Thanks the ability for providing rich information, knowledge graphs (KGs) being incorporated enhance recommendation performance interpretability. To effectively make use of graph, we propose a model in hyperbolic space, which facilitates learning hierarchical structure graphs. Furthermore, attention network is employed determine relative importances neighboring entities certain item. In addition, an adaptive fine-grained regularization mechanism adaptively regularize items their representations. Via comparison using three real-world datasets with state-of-the-art methods, show that proposed outperforms best existing models by 2-16% terms NDCG@K on Top-K recommendation.