The internet has become an indispensable part of our lives, and it provides a platform for enterprises to deliver information about products and services to the customers conveniently. As the amount of this kind of information is increasing rapidly, one great challenge is ensuring that proper content can be delivered quickly to the appropriate customers. Personalized recommendation is a desirable way to improve customer satisfaction and retention . There are mainly three approaches to recommendation engines based on different data analysis methods, i.e., rule-based, content-based and collaborative filtering . Among them, collaborative filtering (CF) requires only data about past user behavior like ratings, and its two main approaches are the neighborhood methods and latent factor models. The neighborhood methods can be user-oriented or item-oriented. They try to find like-minded users or similar items on the basis of co-ratings, and predict based on ratings of the nearest neighbors . Latent factor models try to learn latent factors from the pattern of ratings using techniques like matrix factorization and use the factors to compute the usefulness of items to users. CF has made great success and been proved to perform well in scenarios where user preferences are relatively static. The dynamic nature decides that users’ preferences may drift over time in dynamic recommendation, resulting in different taste to the items in different phases of interest, but it is not well studied in previous studies . In our experiences, the interest cycle differs from user to user, and the pattern how user preferences changes cannot be precisely described by several simple decay functions.