With the explosive growth of open source software, retrieving desired software in open source software communities becomes a great challenge. Tagging open source software is usually a manual process which assigns software with several tags describing its functions and characteristics. Users can search their desired software by matching the keywords. Because of the simplicity and convenience, software retrieval based on tags has been widely used. However, since human effort is expensive and time-consuming, developers are not willing to tag software sufficiently when uploading software projects. Thus automatic software tagging, with tags describing functions and characteristics according to software projects’ text descriptions provided by users, becomes key to effective software retrieval. This article formalizes this problem as a multi-label learning problem and proposes a new multi-label learning method ML-CKNN which can effectively solve this problem when the number of different tags is extremely large. By imposing cost value of wrong classification into multi-label learning, ML-CKNN can effectively solve this imbalanced problem, as each tag instances associated with this tag are much less than those not associated with this tag. Experiments on three open source software community datasets show that ML-CKNN can provide high-quality tags for new uploading open source software while significantly outperforming existing methods.