Understanding Rating Behaviour and Predicting Ratings by Identifying Representative Users

Understanding Rating Behaviour and Predicting Ratings by Identifying Representative Users Rahul Kamath University of Tokyo 7-3-1 Hongo, Bunkyo-ku Tokyo, Japan Masanao Ochi University of Tokyo 7-3-1 Hongo, Bunkyo-ku Tokyo, Japan Yutaka Matsuo University of Tokyo 7-3-1 Hongo, Bunkyo-ku Tokyo, Japan Abstract Online user reviews describing various prod- ucts and services are now abundant on the web. While the information conveyed through review texts and ratings is easily comprehensi- ble, there is a wealth of hidden information in them that is not immediately obvious. In this study, we unlock this hidden value behind user reviews to understand the various dimensions along which users rate products. We learn a set of users that represent each of these dimen- sions and use their ratings to predict product ratings. Specifically, we work with restaurant reviews to identify users whose ratings are in- fluenced by dimensions like ‘Service’, ‘Atmo- sphere’ etc. in order to predict restaurant rat- ings and understand the variation in rating be- haviour across different cuisines. While pre- vious approaches to obtaining product ratings require either a large number of user ratings or a few review texts, we show that it is pos- sible to predict ratings with few user ratings and no review text. Our experiments show that our approach outperforms other conventional methods by 16-27% in terms of RMSE. 1 Introduction With the advent of Web 2.0, a large number of plat- forms including e-commerce sites, discussion fo- rums, blogs etc. have emerged that allow users to express their opinions regarding various businesses, products and services. These opinions are usually in the form of reviews, each consisting of text feedback describing various aspects of the product along with a single numeric rating representing the users’ over- all sentiment about the same (McAuley et al., 2012). Such user review ratings are normally aggregated to provide an overall product rating, which help other people form their own opinion and help them make an informed decision during purchase. However, in case of new products, there is a time delay till a suf- ficient number of ratings that give a ‘complete pic- ture’ of the product can be obtained. In such a sce- nario, the seller of the product may find it useful to identify a few people whose ratings, when combined together, reflect this ‘complete picture’. The seller may then invite these people to review the product and, as a result, reduce the time delay involved in getting the ‘true’ product rating. Review text is unstructured and inherently noisy. But it can be a valuable source of information since users justify their ratings through such text (McAuley and Leskovec, 2013). Users tend to ex- press their sentiments about different aspects of a product in the review text and provide a rating based on some combination of these sentiments (Ganu et al., 2009). However, some users are influenced heavily by one particular aspect of the product and this is reflected in their ratings. For example: While reviewing smartphones, the ratings provided by a user …

Content truncated to 500 words