Multi-label imbalanced classification based on assessments of cost and value

• Published in: Applied intelligence (Dordrecht, Netherlands) Netherlands), 2018-10, Vol.48 (10), p.3577-3590
• Main Authors: Ding, Mengxiao, Yang, Youlong, Lan, Zhiqing
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial Intelligence; Classification; Computer Science; Machines; Manufacturing; Mechanical Engineering; Middle income; Penalty function; Performance enhancement; Preprocessing; Processes
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-018-1156-8

Multi-label imbalanced data comprise data with a disproportionate number of samples in the classes. Traditional classifiers are more suitable for classifying balanced data because the classification performance declines dramatically when the class sizes are imbalanced in multi-label data. In this study, we propose an algorithm that assesses the cost of the majority class and the value of the minority classes to handle the multi-label imbalanced data classification problem. The main idea of our algorithm is to provide a quantitative assessment of the cost of the majority class and the value of the minority class based on an imbalance ratio. In the data preprocessing step, we employ a penalty function to determine the number of majority class instances for elimination. The contributions of an instance determine whether a majority class instance is to be eliminated. In the classification step, we propose a metric to control the cost of the majority class and the value of the minority class. Experiments showed that this algorithm can improve the performance of multi-label imbalanced data classification.