In this work, a semi-supervised hierarchical Bayesian multi-label classifier (SSHBMC) is proposed. SSHBMC is a semi-supervised learning algorithm for hierarchical classification where the hierarchy is a directed acyclic graph and the labels can be associated to multiple paths of labels. SSHBMC builds pseudo-paths-of-labels for each unlabeled instance using information of its nearest labeled instances. A hierarchical Bayesian network classifier, that considers the data distribution while it models the hierarchy, is trained with the labeled and pseudo-labeled data, and later is used to classify new instances based on probabilistic inference. The method was tested in several datasets from functional genomics and compared against related methods, showing in most cases superior performance with statistical significance.