Recommender systems aim at predicting users’ preferences based on abundant information, such as user ratings, demographics, and reviews. Although reviews are sparser than ratings, they provide more detailed and reliable information about users’ true preferences. Currently, reviews are often used to improve the explainability of recommender systems. In this paper, we propose the sentiment based matrix factorization with reliability (SBMF+R) algorithm to leverage reviews for prediction. First, we develop a sentiment analysis approach using a new star-based dictionary construction technique to obtain the sentiment score. Second, we design a user reliability measure that combines user consistency and the feedback on reviews. Third, we incorporate the ratings, reviews, and feedback into a probabilistic matrix factorization framework for prediction. Experiments on eight Amazon datasets demonstrated that SBMF+R is more accurate than state-of-the-art algorithms.

Introduction

Recommender systems are successful in predicting users’ preferences for items using various techniques, such as content-based (Pazzani & Billsus, 2007) and collaborative filtering (CF)-based (Sarwar, Karypis, Konstan, & Riedl, 2001) approaches. Contentbased approaches extract the features of items and build a model or profile of user interests for recommendation. CF-based approaches predict the interests of users by collecting information from similar users or relevant items. One class of popular CFbased methods is the latent factor model, including matrix factorization (MF) (Koren, Bell, & Volinsky, 2009), non-negative MF (Lee & Seung, 2001), and probabilistic MF (PMF) (Mnih & Salakhutdinov, 2008). CF assumes that the ratings can reflect users’ true preferences and items’ real qualities. However, this assumption does not always accord with real-world scenarios for several reasons. First, critical users tend to give a poor rating, whereas tolerant users might give a good rating, which does not necessarily reflect the quality of the item (Raghavan, Gunasekar, & Ghosh, 2012). Second, users who give the same rating may have different degrees of satisfaction (Cheng, Ding, Zhu, & Kankanhalli, 2018; Lloret, Saggion, & Palomar, 2010). Third, rating noise is introduced by fake likes, that is, superior or bad ratings on products presented by a large group of low-paid workers (Mobasher, Burke, Bhaumik, & Williams, 2007). Fourth, a wrong click also has an implicit influence on the unreliability of ratings. Fig. 1 shows some examples of both reviews and ratings selected from Amazon datasets. Here the reviews given by users are inconsistent with their ratings. Fortunately, the increasing information, such as pictures, tags, reviews, and feedback on e-commerce sites, has shed some light on this problem (McAuley, Targett, Shi, & Hengel, 2015; Yu, Zhou, Deng, & Hu, 2018). Most e-commerce sites allow users to provide reviews of products and give feedback (up or down) to evaluate the usefulness of the reviews (Chen, Qi, & Wang, 2012; Connors, Mudambi, & Schuff, 2011; Hart-Davidson, McLeod, Klerkx, & Wojcik, 2010).