The traditional collaborative filtering algorithm does not consider the influence of item popularity in similarity calculation, and the prediction score does not consider the influence of time on the change of user interest, resulting in inaccurate similarity calculation and single recommendation result. To solve these problems, this paper improved the traditional similarity calculation method by combining the item popularity penalty coefficient, improved the recommendation diversity of the algorithm, and integrated the time factor into the prediction method to solve the problem of interest attenuation. Experiments on the 100K and 1M data set of Movielens show that the improved algorithm effectively improves the accuracy and coverage of recommendations.