Collaborative Filtering With Aspect-Based Opinion Mining

In recent years, the Internet has become the greatest source of information that has existed and the electronic retailers provide a huge selection of products. In this context, people are inundated with choices such as what online news to read, what movies to watch, or what products to buy. Recommender systems are tools and techniques that providing suggestions for items to be of use to users according to their particular tastes or interests. Nowadays, more retailers are interested in personalized recommendations because matching the users with the most suitable items is key to enhancing user satisfaction and loyalty. Such systems are especially important to e-commerce websites, so e-commerce leaders like Netflix and Amazon.com have their own recommender systems to improve the user experience.Collaborative filtering (CF) is a widely used technique in recommender systems, which recommend items to a particular user based on other users’ ratings. These ratings, often in the form of a scalar such as1-10stars, represent people’s overall opinions about items. However, the overall ratings cannot provide us more detailed information. For example, a user giving a movie2-star rating may indicate that the user considers this movie to be bad as a whole. However, it is still possible that he likes some aspects (e.g., storyline, music) of the movie very much. On the other hand, most websites enable users to write textual reviews about the items they bought and a lot of research in opinion mining has been conducted to extract subjective opinions from reviews. However, most existing CF methods rely only on users’ overall ratings of items, ignoring the variety of opinions users may have towards different aspects of the items.We therefore propose a new CF framework that integrates multi-faceted opinions in the reviews into the CF process, in order to tap the rich sentiment information embedded in the reviews, and to alleviate the cold start/data sparsity problem. In particular, our framework consists of two components, namely (1) opinion mining, and (2) rating inference. The first component extracts and summarizes the multiple aspects of opinions expressed in the reviews, and generates numerical ratings on the different aspects. The second component uses tensor factorization to infer the overall rating a user may give to an item, forming the basis of item recommendation. The method can be seen as an extension of matrix factorization techniques widely applied in collaborative filtering; however, it can preserve the multi-dimension nature of the data and extract the latent factors in each dimension.To summarize, the proposed approach makes the following contributions.(1)We propose a CF framework that is able to extract fine-grained, multi-faceted opinions from reviews, and integrate them into collaborative filtering.(2)We propose a tensor factorization approach to capture the intrinsic multi-way interactions between users, items, and aspects, and to predict the unknown ratings on items. To the best of our knowledge, we are the first to take this approach. Operating on the tensor composed of the overall and aspect ratings, this approach is able to capture the intrinsic relationships between users, items, and aspects, and provide accurate predictions on unknown ratings.(3)We conduct extensive experiments on a movie data set to verify the effectiveness of our approach, and the experimental results show that our proposal significantly improves the prediction accuracy compared with two baseline methods.