On the advances in pattern recognition using Optimum-Path Forest

Abstract

Pattern recognition (PR) techniques have been paramount to solve different and complex problems in many fields of study. The basic idea behind PR techniques is to compute a model capable of classifying unknown samples. Pattern recognition can be categorized as problems of (i) supervised, and (ii) unsupervised learning. This categorization is related to the existence or absence of labeled data to support the learning process. The learning process is mandatory for PR techniques to learn the data distribution, and the existence of labeled data is an additional information that helps to build more robust models. Many techniques were proposed and are well-established in the literature. The Optimum-Path Forest (OPF) is a graph-based classifier proposed recently, which comprises the models for supervised, semi-supervised and unsupervised learning. The OPF models dataset samples as nodes of a graph and their connections (edges) are defined by some pre-defined adjacency relation. Although very recent, OPF has already been employed in numerous applications and showed promising results, and even outperformed other well-known classifiers. Nonetheless, there is still a lot to be investigated, evaluated and proposed concerning the use and performance of the OPF classifier. This dissertation investigates e proposes variations and modifications to the traditional OPF algorithms concerning supervised and unsupervised learning aiming the assessment of its performance in not yet explored scenarios and to overcome its drawbacks.