Personalized News Recommendation Via Implicit Social Experts

Nowadays, users can easily obtain news information around the world through news site (Google News, Yahoo News) and mobile application (CNN Mobile). However, with the large volume of news available, user can hardly find the news meets their interest efficiently. Therefore, an important research topic is how to help readers automatically find interesting articles match their reading appetites which we call "personalized news recommendation". A variety of news recommendation systems based on different strategies have been proposed to provide news personalization services for online news readers. In this paper, we mainly focus on cold-start and data sparsity problem and propose PRemiSE, a novel personalized news recommendation framework via implicit social experts. This mainly contains the following content:Firstly, how to overcome the data sparsity problem? Many online users read limited news artieles in the entire article repository, and therefore the access matrix is very sparse, and the similarity of users’access patterns can’t be effectively captured.Secondly, how to deal with the cold-start problem? Cold-start problem which including user cold-start and item cold-start refers to how to recommend for new published news or new coming users. Online user groups are evolving and the dynamic nature of news articles makes cold start problem particularly important in this problem.Thirdly, we propose PRemiSE, a hybrid recommendation framework in which the opinions of potential experts on virtual user networks constructed from implicit feedbacks are integrated with nearest neighbors for recommendation. Ratings are generated as the aggregation of user preference, semantic item profiles and preferences of the most influential experts.Empirical results on real world news dataset demonstrate the efficacy of our method. The introduction of news named entities characterize user interest eases the data sparseness problem in a certain extent. Through ensemble the similar user’s taste and expert’s opinion, we effectively mitigate cold start problems. Meanwhile, we analyze the text content of the news makes recommendation results with semantic interpretability.