For improving the performance of the traditional collaborative filtering recommender system, a dynamic user- item-time three-dimensional model based on rolling time windows was proposed, which considers the time seduence problem. I}hen a special collaborative filtering (CF) algorithm was explored to work with the model. 1}he interest scores at different times arc regarded differently according to the time sequence and the similarities between users arc com- posed of components at different times,which increases the timeliness of the algorithm. In addition, the similarities can also be calculated duickly by an incremental formula deduced in this paper so as to improve the scalability of the algo- rithm. At last, some reasonable experiments show that the model and algorithm presented in this paper outperform the traditional 2D collaborative filtering model and algorithm in terms of the hit rate.