Adapted transformer network for news recommendation

Abstract

Online news recommendation aims to provide personalized news for users according to their interests. Existing methods usually learn user preference from their historical reading records in a static and independent way, which ignore the dynamic interaction with the target candidate news. In fact, it is important to fully capture the semantic interaction between user?s historical news and candidate news since the user?s interests would be different in terms of different candidate news. In this paper, we propose a novel news recommendation model with an adapted transformer network. There are three parts in our approach, i.e., a news encoder to learn the semantic features of news, a user encoder to learn the initial representations of users, an adapted transformer module to learn the deep interaction between users and candidate news. The core is that we effectively integrate the historical clicked news and the candidate news into the transformer framework to capture their inherent relatedness. Besides, an additive attention layer is proposed to learn different informativeness of words since different words are differently useful for news or users? representation. We conduct extensive experiments on a real-world dataset from MSN news and the results indicates the effectiveness of our proposed approach.