PGT: news recommendation coalescing personal and global temporal preferences

Given sequential news watch logs of users, how can we accurately recommend news articles? Compared to other items (e.g., movies and e-commerce products) for a recommendation, the worth of news articles decays quickly and massive news articles are published every second. Moreover, people frequently watch popular news articles regardless of their personal tastes to browse remarkable events at a specific time. Current state-of-the-art methods, designed for other target item domains, show low performance when they are used for news recommendation because of these peculiarities of news articles. In this paper, we propose PGT (News Recommendation Coalescing Personal and Global Temporal Preferences), an accurate news recommendation method designed with consideration of the above properties of news articles. PGT sufficiently reflects users’ behaviors by utilizing latent features extracted from both personal and global temporal preferences. Furthermore, we propose an attention-based architecture to extract adequate coalesced features from both of the preferences. We carefully tune each component of PGT to find optimal architecture. Experimental results show that PGT provides the most accurate news recommendation, giving the state-of-the-art accuracy.