A Cost-Sensitive Loss Function for Machine Learning

• Published in: Database Systems for Advanced Applications Vol. 10829; pp. 255 - 268
• Main Authors: Chen, Shihong, Liu, Xiaoqing, Li, Baiqi
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2018; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: IEEM; Loss function; Loss sensitivity; Machine learning
• ISBN: 9783319914541; 3319914545
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-319-91455-8_22

In training machine learning models, loss functions are commonly applied to judge the quality and capability of the models. Traditional loss functions usually neglect the cost-sensitive loss in different intervals, although sensitivity plays an important role for the models. This paper proposes a cost-sensitive loss function based on an interval error evaluation method (IEEM). Using the key points of grade-structured intervals, two methods are proposed to construct the loss function: a piecewise function linking by key points, and a curve function fitting by key points. The proposed function was evaluated against three different loss functions based on a BP neural network. The comparison results show that the proposed loss function based on IEEM made the best prediction of the PM2.5 air quality grade in Guangzhou, China.