Graph-based approaches for pattern recognition techniques are commonly designed for unsupervised and semi-supervised ones. Recently, a novel collection of supervised pattern recognition techniques based on an optimum-path forest (OPF) computation in a feature space induced by graphs were presented: the OPF-based classifiers. They have some advantages with respect to the widely used supervised classifiers: they do not make assumption of shape/separability of the classes and run training phase faster. Actually, there exists two versions of OPF-based classifiers: OPF
(the first one) and OPF
. Here, we introduce a learning algorithm for the last one and we show that a classifier can learns with its own errors without increasing its training set.