Hierarchical Fuzzy Classifier Design Using a Reject Option

Rowan Fuerst, Naoki Masuyama, Yusuke Nojima

Abstract: In the litterature, it has been shown that Multi-objective Fuzzy Genetics-Based Machine Learning can be used to find a range of fuzzy classifiers with varying explainability and accuracy for use in fields such as medicine or finance. It can similarly be used to find a range of fuzzy classifiers considering both the number of features and the accuracy. To increase the classifiers' reliability, a reject option can be used to refuse the classification of uncertain patterns. To reduce the overall classification cost, we would like to use a feature-light classifier most of the time and use a more accurate feature-heavy classifier only for patterns that require a more complex classification boundary, meaning those near the classification boundary. In this paper, we consider the integration of several fuzzy classifiers to combine their individual qualities into a hierarchical fuzzy classifier using a reject option as the condition to relay the pattern between the different classifiers: this way, most input patterns will be classified by the top layer feature-light classifier whose classification cost is low while only the patterns near the classification boundary will be classified by one of the more costly bottom layers feature-heavy classifiers. Through computational experiments, we discuss the performance of the newly formed classifier over several datasets and compare two different reject threshold optimization schemes.