Graph Neural News Recommendation with Unsupervised Preference Disentanglement

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4255–4264 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 4255 Graph Neural News Recommendation with Unsupervised Preference Disentanglement Linmei Hu1, Siyong Xu1, Chen Li1, Cheng Yang1, Chuan Shi1∗ Nan Duan2, Xing Xie2, Ming Zhou2 1School of Computer Science, Beijing University of Posts and Telecommunications, Beijing, China 2Microsoft Research Asia, Beijing, China {hulinmei, xusiyong, leechen, yangcheng, shichuan}@bupt.edu.cn {nanduan, xing.xie, mingzhou}@microsoft.com Abstract With the explosion of news information, per- sonalized news recommendation has become very important for users to quickly find their interested contents. Most existing methods usually learn the representations of users and news from news contents for recommenda- tion. However, they seldom consider high- order connectivity underlying the user-news interactions. Moreover, existing methods failed to disentangle a user’s latent prefer- ence factors which cause her clicks on differ- ent news. In this paper, we model the user- news interactions as a bipartite graph and pro- pose a novel Graph Neural News Recommen- dation model with Unsupervised Preference Disentanglement, named GNUD. Our model can encode high-order relationships into user and news representations by information prop- agation along the graph. Furthermore, the learned representations are disentangled with latent preference factors by a neighborhood routing algorithm, which can enhance expres- siveness and interpretability. A preference reg- ularizer is also designed to force each disentan- gled subspace to independently reflect an iso- lated preference, improving the quality of the disentangled representations. Experimental re- sults on real-world news datasets demonstrate that our proposed model can effectively im- prove the performance of news recommenda- tion and outperform state-of-the-art news rec- ommendation methods. 1 Introduction The amount of news and articles on many news plat- forms, such as Google News1, has been growing ∗Corresponding author:Chuan Shi(shichuan@bupt.edu.cn) 1https://news.google.com/ 𝑢2 𝑢3 𝑢1 𝑑1 𝑑3 𝑑2 𝑑4 ⋯ Layer-2 Layer-1 𝑢2 𝑑1 𝑑4 𝑑2 𝑢1 𝑢1 𝑢3 ？ ⋯ ⋯ ⋯ ⋯ ⋯ ⋯ Figure 1: An illustration of user-news interaction graph and high-order connectivity. The representations of user and news are disentangled with latent preference factors. constantly at an explosive rate, making it difficult for users to seek for news that they are interested in. In order to tackle the problem of information over- load and meet the needs of users, news recommen- dation has been playing an increasingly important role for mining users’ reading interest and provid- ing personalized contents (IJntema et al., 2010; Liu et al., 2010). A core problem in news recommendation is how to learn better representations of users and news. Recently, many deep learning based methods have been proposed to automatically learn informative user and news representations (Okura et al., 2017; Wang et al., 2018). For instance, DKN (Wang et al., 2018) learns knowledge-aware news representation via multi-channel CNN and gets a representation of a user by aggregating her clicked news history with different weights. However, these methods (Wu et al., 2019b; Zhu et al., 2019; An et al., 2019) usu- ally focus on news contents, and seldom consider …

Content truncated to 500 words