Classification with Rejection Option Using the Fuzzy ARTMAP Neural Network

Abstract

The ARTMAP networks are machine learning techniques focused on supervised learning, being known mainly for their ability to learn fast, stable, incremental and online. Despite these advantages, the Fuzzy ARTMAP (FAM) suffers from the categories proliferation problem, leading to a reduction in its performance for unknown samples. Such disadvantage is mainly caused by the overlapping region (noise) between classes. The vast majority of work on this issue has been concerned with alleviating the problem. A technique used to improve the performance of a classifier is the rejection option, which is used to retain the classification of a sample if the decision is not considered to be reliable. Therefore, in this paper, we introduce a variant of the Fuzzy ARTMAP to behave as a classifier with the rejection option. The main idea is to create a region of rejection by looking at the place where the categories proliferate since it occurs precisely in the overlapping region. The proposal was validated by conducting experiments with real datasets, as well as by comparing them with other models (MLP, SVM, and SOM) applied with the same rejection option technique.