Effects of Reject Options using Multiple Thresholds on the Performance of Fuzzy Classifiers

Abstract

Fuzzy genetics-based machine learning can design highly interpretable fuzzy rule-based classifiers. The output of our fuzzy classier is based on the single winner rule with the highest degree of certainty, which is specified as the product of the compatibility grade and rule weight. The final output class is determined without considering the degree of certainty of the second winner rule. Because the degree of certainty of the classification for patterns near the classification boundaries is often low, those patterns tend to be misclassified. To handle this issue, we incorporate a reject option into the classifiers in this paper. We examine three reject options based on a common threshold, class-wise thresholds, and rule-wise thresholds for the difference in the degree of certainty. Experimental results show reject options based on class-wise and rule-wise thresholds can effectively reject classifications with a low degree of certainty.