Learning rule sets and Sugeno integrals for monotonic classification problems

In some variants of the supervised classification setting, the domains of attributes and the set of classes are totally ordered sets. The task of learning a classifier that is nondecreasing w.r.t. each attribute is called monotonic classification. Several kinds of models can be used in this task; in this paper, we focus on decision rules. We propose a method for learning a set of decision rules that optimally fits the training data while favoring short rules over long ones. We give new results on the representation of sets of if-then rules by extensions of Sugeno integrals to distinct attribute domains, where local utility functions are used to map attribute domains to a common totally ordered scale. We study whether such qualitative extensions of Sugeno integral provide compact representations of large sets of decision rules. (C) 2020 Elsevier B.V. All rights reserved.