Study On Key Algorithms In Collaborative Filtering Recommender Systems

As an important means to solve the problem of information overload, Recommender Systems is widely used in various fields such as e-commerce, film community, music community and social networks. In recent years, collaborative filtering becomes the most widely used recommendation techniques due to its simplicity, low data dependency and high prediction accuracy. The key algorithms in collaborative filtering recommender systems include neighborhood-based model and latent factor model. In this thesis, we described the basic principles of neighborhood-based model and latent factor model, analyzed the inadequacies of the existing methods and proposed two new collaborative filtering algorithms, i.e., collaborative filtering based on star users and a modified PMF model incorporating implicit item associations.Star-user-based method is an improved neighbor model. In this method, we firstly train a small set of implicit users, which we call "Star Users", from the original user rating data set. And then we match ordinary users with star users based on specific similarity calculation method. Finally, predictions for the active user are computed as the weighted average of star users’ ratings.The modified probabilistic matrix factorization (PMF) model explores how contextual information influences the relationship between items. First, we decompose the item-context rating matrix and obtain implicit feature vectors of items. Second, build contextualized implicit correlations between items based on these feature vectors. Finally, the implicit correlations between items are integrated into basic PMF model, and we obtain better recommendation results.In addition, we conducted experiments on the Movielens, Netflix, and Yahoo! Music datasets, and compared our proposed methods to conventional methods.