Social bookmarking websites are rapidly growing in popularity. Recommender systems, a promising remedy to the information overload accompanying the explosive growth in content, are designed to identify which unseen content might be of interest to a particular user, based on his or her past preferences. Most previous work in recommendation for social bookmarking suffers from a lack of comparisons between the different available approaches. In this article, we address this issue by comparing and evaluating eight recommendation approaches on four data sets from two domains. We find that approaches that use tag overlap and metadata provide better results for social bookmarking data sets than the transaction patterns that are used traditionally in recommender systems research. In addition, we investigate how to fuse different recommendation approaches to further improve recommendation accuracy. We find that fusing recommendations can indeed produce significant improvements in recommendation accuracy. We also find that it is often better to combine approaches that use different data representations, such as tags and metadata, than to combine approaches that only vary in the algorithms they use. The best results are obtained when both of these aspects of the recommendation task are varied in the fusion process. Our findings can be used to improve the quality of recommendations not only on social bookmarking websites, but conceivably also on websites that offer annotated commercial content.