Learning cost function for graph classification with open-set methods

Abstract

In several pattern recognition problems, effective graph matching is of paramount importance. In this paper, we introduce a novel framework to learn discriminative cost functions. These cost functions are embedded into a graph matching-based classifier. The learning algorithm is based on an open-set recognition approach. An open-set recognition describes a problem formulation in which the training process does not have access to labeled samples of all classes that may show up during the test phase. We also investigate a set of measures to characterize local graph properties. Performed experiments considering widely used datasets demonstrate that our solution leads to better or comparable results to those observed for several state-of-the-art baselines.