Reviews Meet Graphs: Enhancing User and Item Representations for Recommendation with Hierarchical Attentive Graph Neural Network

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4884–4893, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4884 Reviews Meet Graphs: Enhancing User and Item Representations for Recommendation with Hierarchical Attentive Graph Neural Network Chuhan Wu1, Fangzhao Wu2, Tao Qi1, Suyu Ge1, Yongfeng Huang1,and Xing Xie2 1Department of Electronic Engineering, Tsinghua University, Beijing 100084, China 2Microsoft Research Asia, Beijing 100080, China {wu-ch19, qit16, gsy17, yfhuang}@mails.tsinghua.edu.cn, {fangzwu, xing.xie}@microsoft.com Abstract User and item representation learning is crit- ical for recommendation. Many of exist- ing recommendation methods learn represen- tations of users and items based on their rat- ings and reviews. However, the user-user and item-item relatedness are usually not consid- ered in these methods, which may be insuffi- cient. In this paper, we propose a neural rec- ommendation approach which can utilize use- ful information from both review content and user-item graphs. Since reviews and graphs have different characteristics, we propose to use a multi-view learning framework to in- corporate them as different views. In the re- view content-view, we propose to use a hier- archical model to first learn sentence repre- sentations from words, then learn review rep- resentations from sentences, and finally learn user/item representations from reviews. In addition, we propose to incorporate a three- level attention network into this view to se- lect important words, sentences and reviews for learning informative user and item repre- sentations. In the graph-view, we propose a hierarchical graph neural network to jointly model the user-item, user-user and item-item relatedness by capturing the first- and second- order interactions between users and items in the user-item graph. In addition, we apply at- tention mechanism to model the importance of these interactions to learn informative user and item representations. Extensive experiments on four benchmark datasets validate the effec- tiveness of our approach. 1 Introduction Precisely learning user and item representations is critical for recommendation (Tay et al., 2018). Many existing recommendation methods learn user and item representations from their rating ma- trix (Koren et al., 2009; Mnih and Salakhutdinov, 2008; Berg et al., 2017). For example, Koren et I feel this is a must read for any Star Wars fan of the prequels and original trilogy. The story is very well written and makes it hard to put this book down. I loved this book. I believe this is the 17th Star Wars book I've read. I've read from the from the earliest days, before the old republic up until Rogue One and cover to cover, this is hands down the best book I've read thus far. User-1 User-2 Item-1 Item-2 Figure 1: Two example users and items. al. (2009) proposed to use singular value decom- position (SVD) to learn latent user and item rep- resentations based on the review ratings that users gave to items. However, there are massive users and items in online platforms, making the rating matrix between users and items very sparse. Thus, it …

Content truncated to 500 words