Combinação de classificadores baseados em floresta de caminhos ótimos

Abstract

Machine learning techniques have been actively pursued in the last years, mainly due to the great number of applications that make use of some sort of intelligent mechanism for decision-making processes. In this context, among the several studies on classification techniques and how to improve them, the ensemble of classifiers has achieved considerable evidence in the literature. In this circumstance, a classifier with significant growth is the technique called Optimum-Path Forest (OPF), which is considerable ease to manipulate, has no parameters in some versions, and it is efficient in the training phase. Since OPF is a relatively new technique in the literature, and we have few studies on ensemble of OPF classifiers only, this work aims to provide a more detailed study in ensemble techniques regarding the OPF classifier. This work has proposed an improved version of OPF, which learns a score-based confidence level for each training sample in order to turn the classification process “smarter” (i.e., more reliable), which is further used in a combination process with majority voting. Furthermore, we also proposed the combination of classifiers using an ensemble pruning strategy driven by meta-heuristics based on quaternions. In addition, we proposed an extension of the ensemble pruning using OPF classifiers in the context of remote sensing images. Finally, the probabilistic OPF was proposed, since the OPF presents only abstract outputs. Experimental results over synthetic and real datasets showed the effectiveness and efficiency of the proposed approaches for classification problems.