Enriching user and item profiles for collaborative filtering : from concept hierarchies to user-generated reviews

Abstract

Collaborative Filtering (CF) is a recommender systems technique that generates personalized recommendations for users based on user preferences. Such preferences are usually expressed in the form of numerical ratings, or binary votes such as purchase data. Despite its considerable success and popularity in both research and practice, CF suffers from the problems of data sparseness and cold-start recommendation, which is an extreme form of data sparseness. Specifically, CF algorithms have difficulty with generating reliable recommendations when data are sparse, and they cannot recommend items that have not received any ratings from users. This thesis addresses the problems of data sparseness and cold-start recommendation of CF along two dimensions. Firstly, we developed two novel recommendation algorithms based on association rule mining techniques. The proposed algorithms, namely FARAMS and CLARE, exploit the relationships between items that are encoded in the concept hierarchies of the items when users' preference data are too limited for generating recommendations. Specifically, FARAMS makes use of interesting associations between item categories to find recommendable items for users having limited known preferences, while CLARE generates recommendations for a given cold-start item by finding other items in the system that are highly correlated with the attributes of the cold-start item. We evaluated both algorithms based on widely adopted benchmarking datasets of CF. Results show that both algorithms outperform related algorithms in addressing data sparseness and the cold-start problem under similar experimental settings. Secondly, we investigated the use of user-generated reviews for generating personalized recommendations. We made three major contributions in this area. First, we collected and analyzed a set of movie reviews to understand how user opinions are expressed in user-generated reviews, which are free-form texts written in natural language. Based on the results of our analysis, we proposed a novel method for determining the sentimental orientations and strength of user opinions. Second, we proposed a rating inference framework, namely PREF, for augmenting ratings for CF. PREF aims at determining and representing the overall sentiments expressed in reviews as numerical ratings that can readily be used by existing CF algorithms. In other words, PREF enables existing CF algorithms to utilize textual reviews as an additional source of user preferences, thereby lessens the problem of data sparseness. Third, we found that user-generated reviews contain valuable information for constructing the interest profiles of users and domain items based on a real-world dataset of tourist attraction reviews. Using such information for generating personalized recommendations significantly improve the prediction quality and coverage of traditional CF algorithms. While existing CF algorithms operate on numerical ratings or binary votes of items, our research represents an important pioneering step towards a novel CF paradigm based on user-generated reviews.