DACNN

Abstract

Recently, news recommendation has attracted great attentions as it helps users quickly find news satisfying their preferences. It is important to note that both the interests of users and the news readers change over time. Although existing news recommendation methods have achieved promising performance, they treat the representation of news as static and ignore the dynamic nature of news, i.e. news attracts different users at different times and thus may have different features. In this paper, we propose the Dynamic Attentive Convolution Neural Network (DACNN) to solve the above issues. Specifically, we extract features from the clicked history of news as dynamic representations of news and from the browsed history of users as dynamic representations of users. Moreover, we propose to employ a shared CNN with inner-attention to learn user-item interactions from the dynamic representation. Extensive experiments are conducted on two real-world datasets and have proved the superiority of our model.