Flexible Multi-Class Cost-Sensitive Thresholding

Abstract

Classification address the categorization of input data into predefined classes or categories based on their characteristics. Thresholding techniques predict the optimal class for an observation given a score and a missclassification error cost specification. In multi-class classification, existing algorithms assume that a score is available for each of the possible responses. However, there are scenarios where multiple classes can be predicted from a single score. Based on the good performance and flexibility of the 2-DDR algorithm proposed in C-Rella et al. (2024), in this paper we extend it to the multi-class setting. With this approach, the optimal multi-class labeling is obtained from a single score, a problem not previously studied. Furthermore, a more efficient version of the original algorithm is proposed. The good performance of the proposed technique is demonstrated in an extensive simulation study and on four real data sets.

Keywords: Cost-sensitive classification, thresholding, multi-class classification, decision making