User-generated content based recommendation systems for investment and e-commerce

Abstract

Nowadays, user-generated content based recommendation systems (UGC-Recsys) have become a very popular trend in recent years. In this thesis, recommendation systems based on three types of user-generated data are proposed, and their efficient evaluation is studied.

First, this thesis focuses on a very useful user-generated content in investment domain, i.e., online investor opinions. The posts of investors who publish their investment views and suggestions are very useful for users who make investing decisions according to them. However, these online investor opinions vary greatly in quality and are of large scales. Thus, this thesis studies how to extract high-quality investor opinions from a massive collection of investor opinions posted on the web and make use of them for improving investment recommendation tasks. Our experiments on real-world datasets will show that the approaches presented in this thesis could recommend investor opinions and stocks that would bring users substantial profits.

Our attention is then turned to product reviews that contain users’ viewpoints about products. Our goal is to improve recommendation of product reviews, by using personalization criteria. This is motivated by the fact that the importance of product aspects to different users may vary and users prefer to focus on the most important aspects to them. Previous work on ranking reviews based on quality and coverage is improved in the thesis to also consider the personalized preferences of users on product aspects. An experimental evaluation with two public review datasets demonstrates the effectiveness of our approach on recommending reviews that have high quality, coverage, and relevance to the aspects that are important for the user.

A wide range of user-generated content exists in the e-commerce domain. Product reviews are one kind of unstructured user-generated content while there are also several kinds of more structured user-generated content. For example, peoples could share their friend lists (social connections) or their check-in records (geographical locations) in many location based social network (LBSN) websites such as Foursquare and Gowalla. This thesis focuses on social connections and check-in records obtained from LBSN platforms. For convenience, we call the collection of social connections and check-in records from LBSN platforms as LBSN data. Based on LBSN data, location recommendation becomes a typical function of location service based e-commerce sites. In this thesis, we focus on the social-activity related locations (briefly denoted as activity locations) which indicate the locations could be mapped to real-world social activities. Our work is the first one to consider an important characteristic of activity locations: people like to go to these activity locations with their friends together, and improves the effectiveness of recommendation on activity locations in two directions. First, the problem of location-partner recommendation (i.e., for each recommended activity location, find a suitable partner for the user) is studied. Second, assuming that users tend to select the activities for which they can find suitable partners, a partner-aware activity location recommendation model is proposed. Finally, experiments on real data are conducted, which verify the effectiveness of location-partner recommendation and partner aware activity-location recommendation.