Recommend system is an effective method for people to get useful knowledge from mass information. It has attracted widespread attention in both academia and industry. Collaborative filtering (CF) is the most popular algorithm in the research of Recommend system. However most of current CF algorithms are static models, which do not take into account of user interest changing. The paper proposed a hybrid recommend method, which capture user's long-term interests with Gaussian probabilistic latent semantic (PLSA) algorithm, at the same time, capture user's short-time interests with rating window. The experimental results obtained on Movielens dataset and Netflix dataset clearly show that the new algorithm is more accurate than traditional user-based algorithm and PLSA algorithm.