Learning to weight similarity measures with Siamese networks: A case study on optimum-path forest

Abstract

Recent advances in machine learning algorithms have been aiding humans and improving their decision-making capacities in various applications, such as medical imaging, image classification and reconstruction, object recognition, and text categorization. A graph-based classifier, known as Optimum-Path Forest (OPF), has been extensively researched in the last years, mainly due to its parameterless nature and state-of-the-art results compared to well-known literature classifiers, for example, support vector machines. Nevertheless, one drawback concerning such an approach lies in its distance calculation, which has to be selected from a range of formulae and computed between all nodes to weigh the graph's arcs, and hence time-consuming. Therefore in this work, we propose to address such a problem by precomputing the arcs' distances through a similarity measure obtained from Siamese networks. The idea is to employ the same training set used by the OPF classifier to train a Siamese network and calculate the samples' distance through a similarity measure. The experimental results show that the proposed method is suitable, where the similarity-based OPF achieved comparable results to its standard counterpart and even surpassed it in some datasets. Additionally, the precalculated similarity matrix lessens the burden of recalculating the distances for every new classification. © 2022 Copyright