Although various algorithms for multi-label classification have been developed in recent years, there is little, if any, information as to when each method is beneficial. The main goal of this paper is to compare the classification performance of several multi-label algorithms and to develop a set of rules or tools that will help in selecting the optimal algorithm according to a specific dataset and target evaluation measure. We utilize a meta-learning approach allowing fast automatic selection of the most appropriate algorithm for an unseen dataset based on its descriptive characteristics. We also define a list of characteristics specific for multi-label datasets. The experimental results indicate the applicability and usefulness of the meta-learning approach.