Semi-supervised Pattern Classification Using Optimum-Path Forest

Abstract

We introduce a semi-supervised pattern classification approach based on the optimum-path forest (OPF) methodology. The method transforms the training set into a graph, finds prototypes in all classes among labeled training nodes, as in the original supervised OPF training, and propagates the class of each prototype to its most closely connected samples among the remaining labeled and unlabeled nodes of the graph. The classifier is an optimum-path forest rooted at those prototypes and the class of a new sample is determined, in an incremental way, as the class of its most closely connected prototype. We compare it with the supervised version using different learning strategies and an efficient method, Transductive Support Vector Machines (TSVM), on several datasets. Experimental results show the semi-supervised approach advantages in accuracy with statistical significance over the supervised method and TSVM. We also show the gain in accuracy of semi-supervised approach when more representative samples are selected for the training set.