Accuracy-Rejection Tradeoff Analysis in Fuzzy Classifiers with a Two-Stage Reject Option

Abstract

In general, fuzzy classifiers have high interpretability because fuzzy classifiers can linguistically explain the classification reason by fuzzy sets used in the antecedent conditions of rules. A reject option that rejects patterns near the boundaries between different classes is an approach to increase the reliability of fuzzy classifiers. However, the conventional threshold-based reject option may reject more patterns than necessary to achieve high reliability. In this paper, we propose a two-stage reject option where after the threshold-based decision, the k-nearest neighbor is used for patterns with low confidence value than the threshold. If the class labels predicted by the k-nearest neighbor and the fuzzy classifier are the same, the fuzzy classifier outputs the predicted class label without rejection. Through computational experiments, we discuss the relationship between accuracy and the rejection rate.