Despite of success in both research and industry, traditional collaborative filtering (CF) based recommender systems suffer from a fundamental problem, which lies in its dependence on users' numeric ratings as its sole source of user preference information. User ratings are often unable to fully represent user preferences. As a result, it is difficult and error prone to identify genuinely similar users based on ratings only. On the other hand, online consumer product reviews have become a common source for consumers to share and acquire information about products, but there have been very few studies on how those text reviews can be analyzed and integrated with traditional CF approaches to improve the prediction of consumers' preferences. We propose a novel approach to memory-based collaborative filtering called urCF (User Review enhanced Collaborative Filtering) that integrates user text reviews and user numeric ratings in order to model users' preferences better and in turn improve the performance of CF-based recommender systems. This research extracts user opinions on individual item features from online reviews, and proposes a new weighting scheme by following the general idea of TF-IDF to measure the priority of item features in influencing users' overall opinions on different items. This study also explores and compares two different methods for integrating user opinion into user similarity measurement. The proposed urCF system is evaluated against existing approaches using a dataset collected from Yahoo! Movies. The results show that urCF significantly improves the performance of memory-based CF systems.
展开▼
