Visual Approach to Support Analysis of Optimum-Path Forest Classifier

Abstract

Optimum-path forest (OPF) is a graph based classifier in which the training process computes optimum-path trees rooted by prototype instances. Thus, one or more optimum-path trees represent each class and the testing process is based on identifying which optimum-path tree would contain a test sample. Usually, OPF performance is analyzed based on measures computed from training and testing process, such as f-score and correct classification rate (accuracy). This paper proposes an approach based on visualization to support understanding of OPF training and testing processes. The visual approach uses multidimensional projection techniques to reduce the feature space dimensionality and to generate graphical representation from instances similarities. As a result, one can visualize, analyze and understand each step of OPF classifier: generation of the minimum-spanning tree, prototypes choosing, computation of optimum-path trees, and test samples classification. The experiments show that our approach is useful to understand how the prototypes are chosen, to identify what are the best prototypes, to visualize how the training dataset size influences the OPF performance, to analyze how a weak feature space can impact the OPF performance, and to identify some insights about OPF classifier as a whole.