A context-aware matrix factorization recommender algorithm

Latent factor models based on matrix/tensor factorization techniques are widely recognized in recommender systems. However, matrix/tensor factorization techniques are computationally intensive, which greatly limits their usages in real world projects. To address this problem, we improve the classic matrix factorization models by establishing fuzzy mapping relationships between contexts and latent factors, and making use of contextual information to initialize user/item feature vectors. Experimental results on standard test set MovieLens 1M show that the proposed algorithms could achieve far better prediction accuracy while reducing the iteration number by 25%.