Hierarchical User and Item Representation with Three-Tier Attention for Recommendation

Proceedings of NAACL-HLT 2019, pages 1818–1826 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1818 Hierarchical User and Item Representation with Three-Tier Attention for Recommendation Chuhan Wu1, Fangzhao Wu2, Junxin Liu1, and Yongfeng Huang1 1Department of Electronic Engineering, Tsinghua University Beijing 100084, China 2Microsoft Research Asia {wuch15,ljx16,yfhuang}@mails.tsinghua.edu.cn wufangzhao@gmail.com Abstract Utilizing reviews to learn user and item repre- sentations is useful for recommender systems. Existing methods usually merge all reviews from the same user or for the same item into a long document. However, different reviews, sentences and even words usually have dif- ferent informativeness for modeling users and items. In this paper, we propose a hierarchical user and item representation model with three- tier attention to learn user and item representa- tions from reviews for recommendation. Our model contains three major components, i.e., a sentence encoder to learn sentence represen- tations from words, a review encoder to learn review representations from sentences, and a user/item encoder to learn user/item represen- tations from reviews. In addition, we incorpo- rate a three-tier attention network in our model to select important words, sentences and re- views. Besides, we combine the user and item representations learned from the reviews with user and item embeddings based on IDs as the final representations to capture the latent fac- tors of individual users and items. Extensive experiments on four benchmark datasets vali- date the effectiveness of our approach. 1 Introduction Learning accurate user and item representations is very important for recommender systems (Tay et al., 2018). Many of existing recommendation methods learn user and item representations based on the ratings that users gave to items (Koren et al., 2009; Mnih and Salakhutdinov, 2008). For exam- ple, Koren et al. (2009) proposed a matrix factor- ization method based on SVD to learn latent rep- resentations of users and items from the rating ma- trix between users and items. However, since the numbers of users and items in online platforms are usually huge, and the rating matrix between users and items is usually very sparse, it is quite diffi- ＊＊＊＊合 Defrag and cleanup, then you have a great laptop! July 4, 2018 Style: Laptop Only Verified Purchase I bought this laptop yesterday. This is a great laptop if you immediately run maintenance checks on it (defrag, disk cleanup} and remove a little bloatware. It is not a laptop to game with, but as a working/school laptop, you're getting a great bang for your buck. Only giving it four stars just because of the above mentioned things I did afterward, but by no means is this a ho 「rible laptop. 106 people found this helpful ***** MULTIMEDIA LAPTOP June 22, 2018 Style: Laptop Only Verified Purchase This Laptop Great! One person found this helpful Figure 1: Two example reviews. cult for those rating based recommendation meth- ods to learn accurate user and item representa- tions (Zheng et al., 2017; Tay et al., 2018). Luckily, in many online platforms such as Ama- zon and IMDB, there …

Content truncated to 500 words