A multi-task learning approach in hyperbolic space for Knowledge-Enhanced Recommendation

• 1、School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044

Abstract：Recommender systems aim to recommend personalized online content and information for users, effectively alleviating the information overload problem. In recent years, a series of works have found that the rich semantic information in knowledge graphs can benefit collaborative filtering methods. Existing knowledge graph-based recommender systems learn entity representations in Euclidean space, which cannot adequately capture the hierarchical structural information and complex logical relationships embedded in the collaborative knowledge graph. To this end, this paper proposes a multi-task approach in hyperbolic space for knowledge-enhanced recommendation (MHKR). MHKR initializes embeddings in a predefined three-level hyperbolic space, then optimizes embeddings through a personalized recommendation module and a knowledge graph embedding module, and finally trains the two tasks jointly based on a multi-task learning framework. Empirical results on three benchmark datasets show that MHKR outperforms state-of-the-art methods.

Keywords： Recommendation System；Knowledge Graph；Hyperbolic Geometry；Metric Learning；Multi-task Learning