Dynamic news recommendation with hierarchical attention network

News recommendation is an effective information dissemination solution in modern society. In general, news articles can be modeled from multiple granularities: sentence-, element-and news-level. However, the first two levels have been largely ignored in existing methods and it is also unclear how such multi-granularity modeling can enhance news recommendation. In this paper, we propose a novel dynamic model for news recommendation. A unique perspective of our model is to discriminate the contributions of previously interacted contents for triggering the next news-reading, in sentence-, element-and news-level simultaneously. To this end, we design a hierarchical attention network of which the lower layer learns the impacts of sentences and elements, while the upper layer captures disparity of news. Moreover, we incorporate a time-decaying factor to reflect the dynamism, as well as convolution neural networks for learning sequential influence. Using three real-world datasets, we conduct extensive experiments to verify the superiority of our model, compared with several state-of-the-art approaches.