News Recommendation with Knowledge Graph and News Popularity

Abstract:
Knowledge Graphs (KG) have shown great success in recommendation. This is attributed to the rich attribute information contained in KG to solve data sparsity problem and improve item and user representations as side information. However, existing knowledge-aware news recommendation methods didn't consider the contextual information of knowledge entities. In this paper, we propose KGNP, which obtains the contextual information of the entity such as frequency and position through a context embedding layer. In addition, we propose to incorporate news popularity as the contextual information about news to alleviate differentiation and the cold-start differentiation problems of personalized news recommendation. Then a Long Short Term Memory network is applied to the user embeddings by user's history click sequences. Experimental results demonstrate that KGNP can improve the accuracy of news recommendation.