User-based collaborative filtering is an effective and widely-used method in recommender systems. But the data sparsity (the ratings or actions are very sparse for resources) is an inherent limitation of this method. In order to solve the data sparsity, an approach which uses K-medoids algorithm in collaborative filtering is proposed. And the content features of resources are applied to clustering. This approach mainly includes three parts. Firstly, the resources are clustered by K-medoids algorithm. Secondly, the user-behavior data are condensed based on the clustered resources. Thirdly, the recommended list is generated via user-based collaborative algorithm using the compressed user-behavior data. Finally, experiments on data from an Internet education resources sharing platform indicate that the proposed method brings significant improvement both on Recall and Precision in sparse dataset.