Combining User Specific and Global News Features for Neural News Recommendation.

News recommendation engines play a vital role in online news services. Such systems help users discover news articles they may want to read, thus increasing user experience and engagement. Existing news recommendation methods usually model users' interest from their reading history by learning user and news representations. Those methods represent news by features specific to users to create personalized user interest models and recommend news similar to user interest. However, in many cases, users may also read the news that is not in their usual interest, for example, breaking news. Such news often has global characteristics that are independent of specific users. In this paper, we propose a neural news recommendation method that combines user-specific news models with global news models to better deal with those situations. To create informative news models, we adopt an attentive neural approach that can learn representations of the news from different news aspects such as title, category, and news contents. The attentive mechanism allows the model to focus on an aspect that attracts global interest from different users. Experiments on a real-world dataset show that our model achieves improved performance over the state-of-the-art news recommendation method and is able to recommend candidate news of both types: those are relevant to a specific user and those are of common interest for different users.