Next News Recommendation via Knowledge-Aware Sequential Model

A news recommendation system aims to predict the next news based on users' interaction histories. In general, the clicking sequences from the interaction histories indicate users' latent preference, which plays an important role in predicting their future interest. Besides, news articles consist of considerable knowledge entities which have deep connections from common sense of human. In this paper, we propose a Self-Attention Sequential Knowledge-aware Recommendation (Saskr) system consisting of sequential-aware and knowledge-aware modelling. We use the self-attention mechanism to uncover sequential patterns in the sequential-aware modelling. The knowledge-aware modelling leverage the knowledge graph as side information to mine deep connections between news, thus improving diversity and extensibility of recommendation. Content-based news embeddings help to address the item cold-start problem. Through extensive experiments on the real-world news dataset, we demonstrate that the proposed model outperforms state-of-the-art deep neural sequential recommendation systems.