Recommender systems have become extremely popular in recent years since they can provide personalized information to user from a large amount of data, which is typically noisy and hard to exploit. Traditional approaches mainly leverage the user-item rating matrix forrecommendation. Beyond the rating matrix, however, there exists rich side information in recommender systems, which is a good source to improve the performance of rating prediction. In this thesis, we studied three types of side information (i.e., content, temporal, spatial), pointed out some open issues that are unsolved by the existing models and proposed our solutions in these areas.We incorporate side information with some domain knowledge to improve the recommender systems. In recommendation with content information, we proposed a feature-centric model to analyze the feature-level preferences instead of the item-level preferences, thus, make prediction according to feature-level preferences. We further proposed a recommendation by blending content and attributes in heterogeneous networks. In recommendation with temporal information, we proposed temporal matrix factorization to model the user’s interest shift over time; such changes are essential for developing accurate recommender systems. In recommendation with spatial information, we proposed a cross-region collaborative filtering method to deal with the POI (Point of Interest) recommendation when the user travels to a new place; in this model, the long-term and short-term preferences are considered respectively. All these models are evaluated in real life data sets with the state-of-the-art methods.
展开▼
