Score-Oriented Loss (SOL) functions. (December 2022)

• Record Type: Journal Article
• Authors: Marchetti, F.; Guastavino, S.; Piana, M.; Campi, C.
• Is Part Of: Pattern recognition. Volume 132(2022)
• Publication Date: 2022-12
• Subjects: Supervised machine learning -- Binary classification -- Loss functions -- Skill scores; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108913
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: In deep learning loss minimization and score maximization are intertwined issues. SOL functions guarantee an a priori optimization of a given skill score. SOL perspective is to treat the score-related threshold as a random variable. The optimal threshold is driven by the density map of the a priori distribution. Classification tests confirm the automatic threshold optimization provided by SOLs. Abstract: Loss functions engineering and the assessment of prediction performances are two crucial and intertwined aspects of supervised machine learning. This paper focuses on binary classification to introduce a class of loss functions that are defined on probabilistic confusion matrices and that allow an automatic and a priori maximization of the skill scores. These loss functions are tested in various classification experiments, which show that the probability distribution function associated with the confusion matrices significantly impacts the outcome of the score maximization process, and that the proposed functions are competitive with other state-of-the-art probabilistic losses.