Collaborative filtering have become increasingly important with the development of Web 2.0. Online shopping service providers aim to provide users with quality list of recommended items that will enhance user satisfaction and loyalty. Matrix factorization approaches have become the dominant method as they can reduce the dimension of the data set and alleviate the sparsity problem. However, matrix factorization approaches are limited because they depict each user as one preference vector. In practice, we observe that users may have different preferences when purchasing different subsets of items, and the periods between purchases also vary from one user to another. In this work, we propose a probabilistic approach to learn latent clusters in the large user-item matrix, and incorporate temporal information into the recommendation process. Experimental results on a real world dataset demonstrate that our approach significantly improves the conversion rate, precision and recall of state-of-the-art methods.