Collaborative filtering algorithm based on structured projective nonnegative matrix factorization

In collaborative filtering algorithm, the classical matrix factorization may destroy the adjacent structures among data points from high dimension to low dimension. A novel collaborative filtering algorithm based on structured projective nonnegative matrix factorization (CF-SPNMF) was proposed in order to overcome the problem. The algorithm contains both offline learning and online searching. In the offline learning stage, projective nonnegative matrix factorization was applied to obtain the low dimensional latent factors of user preference without changing the intrinsic structure of users' cluster. In the online searching stage, cosine similarity was used to measure the similarity between the target user and training users based on the latent factors inferred in the offline stage. Then the most similar neighbor set was further found. The extensive experiments on real-world data set demonstrate that the proposed CF-SPNMF achieves better rating prediction performance than traditional methods using either matrix factorization or users clustering in original rating matrix. ©, 2015, Zhejiang University. All right reserved.