The Multiclass ROC Front method for cost-sensitive classification. (April 2016)

• Record Type: Journal Article
• Authors: Bernard, Simon; Chatelain, Clément; Adam, Sébastien; Sabourin, Robert
• Is Part Of: Pattern recognition. Volume 52(2016:Apr.)
• Page Start: 46
• Page End: 60
• Publication Date: 2016-04
• Subjects: Multiclass classification -- Cost-sensitive classification -- ROC optimization -- Multi-objective optimization -- SVM classifier; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2015.10.010
• Languages: English
• ISSNs: 0031-3203

Abstract

Abstract: This paper addresses the problem of learning a multiclass classification system that can suit to any environment. By that we mean that particular (imbalanced) misclassification costs are taken into account by the classifier for predictions. However, these costs are not well known during the learning phase in most cases, or may evolve afterwards. There is a need in that case to learn a classifier that can potentially suit to any of these costs in prediction phase. The learning method proposed in this work, named the Multiclass ROC Front (MROCF) method, responds to this issue by exploiting ROC-based tools through a multiobjective optimization process. While this type of ROC-based multiobjective optimization approach has been successfully used for two-class problems, it has never been proposed in real-world multiclass classification problems. Experiments led on several real-world datasets show that the MROCF method offers a major improvement over a cost-insensitive classifier and is competitive with the state-of-the-art cost-sensitive optimization method on all but one of the 20 datasets. Abstract : Highlights: We propose a new method for multiclass cost-sensitive classification when misclassification costs are unknown during training. It is based on a multi-model approach and can suit to any cost-sensitive environment in prediction. It makes use of ROC-based multi-objective optimization algorithms. The method is compared to a cost-insensitive method and a state-of-the-art cost-sensitive optimization method. It outperforms both methods for most of the datasets tested.