In order to solve the problems of poor scalability and low recommendation accuracy of traditional item similarity measurement methods, an improved collaborative filtering recommendation algorithm based on item feature clustering was proposed. Based on full consideration of the similarity of item scores, clustering methods are used to cluster similar items into sets to more effectively discover similar neighbor items of the target item. Based on the MovieLens data set, the experimental results show that the recommendation results of the algorithm are significantly better than the current commonly used collaborative filtering methods in terms of prediction effectiveness, prediction error and recommendation accuracy. The algorithm only considers the ratio of item rating values and does not fully utilize the absolute rating values of items. Comprehensive comparison, the algorithm has obvious improvement in scalability and accuracy compared with the traditional method.