Dynamic Hyperbolic Embeddings with Graph-Centralized Regularization for Recommender Systems

抄録

In this work, we propose two techniques for accurate and efficient hyperbolic embeddings for real-world recommender systems. The first technique is regularization. We found that the graphs of various recommendation datasets exhibit hierarchical or tree-like structures suitable for hyperbolic embeddings, while these structures are not well modeled by the original hyperbolic embeddings. Hence, we introduce a regularization term in the objective function of the hyperbolic embeddings for forcibly reflecting hierarchical or tree-like structures. The second technique is an efficient embedding method, which only updates the embedding of items that are recently added in a recommender system. In an offline evaluation with various recommendation datasets, we found that the regularization enforcing hierarchical or tree-like structures improved HR@10 up to +9% compared to hyperbolic embeddings without the regularization. Moreover, the evaluation result showed that our model update technique could achieve not only greater efficiency but also more robustness. Finally, we applied our proposed techniques to a million-scale news recommendation service and conducted an A/B test, which demonstrated that even 10-dimension hyperbolic embeddings successfully increased the number of clicks by +3.7% and dwell time by +10%.