This paper presents a classifier which uses a tree-based Neural Network (NN), and uses both, unlabeled and labeled instances. First, we learn the structure of the data distribution in an unsupervised manner. After convergence, and once labeled data become available, our strategy tags each of the clusters according to the evidence provided by the instances. Unlike other neighborhood-based schemes, our classifier uses only a small set of representatives whose cardinafity can be much smaller than that of the input set. Our experiments show that, on average, the accuracy of such classifier is reasonably comparable to those obtained by some of the state-of-the-art classification schemes that only use labeled instances during the training phase. The experiments also show that improved levels of accuracy can be obtained by imposing trees with a larger number of nodes.