Author(s)

Hongru Sun, Wancheng Ni and Yiping Yang

Corresponding Author

Hongru Sun

ABSTRACT

Knowledge graphs (KGs), which provide rich semantic information, have proven to be effective in alleviating sparsity and cold start problems in recommender systems. Most existing KG-based methods concentrate on modeling relationship between users and items with constant item-attribute triplets, but neglect the particularity of the KG in the recomme-ndation scenario, that is, the item-attribute triplets should have different importance. In this work, we propose a hierarchical knowledge-based recommendation model (HKRM), which exploits the item-attribute hierarchy and jointly learn knowledge representation with recom-mender system. By measure the proximity between items and attributes, we obtain two dist-ributions: user preference and the candidate item's attributes importance. Then these two di-stributions are respectively transformed into the user representation and item representation, which are used in the recommender system in a duet matching way. To eliminate the doma-in difference, we propose a hypothesis based on distribution distance and a joint learning m-ethod to guide the learning of knowledge representation. We conduct extensive experiments on two datasets related to movies and books. The results demonstrate a significant improv-ement over the state-of-the-art baselines.

KEYWORDS

User fine-grained preference; domain difference; hierarchical knowledge-based recommendation