Towards Better User Preference Learning for Recommender Systems

In recent years, recommender systems have become widely utilized by businesses across industries. Given a set of users, items, and observed user-item interactions, these systems learn user preferences by collective intelligence, and deliver proper items under various contexts to improve user engagements and merchant profits. Collaborative Filtering is the most popular method for recommender systems. The principal idea of Collaborative Filtering is that users might be interested in the items that are preferred by users with similar preferences. Therefore, learning user preferences is the core technique of Collaborative Filtering. In this thesis, we study new methods to help us better understand user preferences from three perspectives. We first dive into each rating that users give on the items, and study the reasons behind the ratings by analyzing user reviews. We propose a unified model that combines the advantages of aspect-based opinion mining and collaborative filtering, which could extract latent aspects and sentiments from reviews and learn users' preferences of different aspects of items collectively. In our next work, we study the problem from each user's perspective, and propose a general and flexible model that embraces several popular models as special cases. The new model achieves better top-N recommendation results on several popular data sets. Finally, we study how to utilize the general structure of the user-item matrix to better apply collaborative filtering methods. We propose a co-clustering based method that first partitions the users and items into several overlapping and focused subgroups, and then applies collaborative filtering methods within each subgroup. The final recommendations for users are aggregated from the results from the subgroups they are involved in. Experimental results show that this method could produce better recommendations than other co-clustering methods and methods that directly apply collaborative filtering on the original user-item matrix.