Multi-label learning algorithm with SVM based association

• Published in: 高技术通讯（英文版） Vol. 25; no. 1; pp. 97 - 104
• Main Authors: Feng Pan, Qin Danyang, Ji Ping, Ma Jingya, Zhang Yan, Yang Songxiang
• Format: Journal Article
• Language: English
• Published: Key Laboratory of Electronic and Communication Engineering, Heilongjiang University, Harbin 150080, P. R. China%School of Electronics and Information Engineering, Harbin Institute of Technology, Harbin 150001, P. R. China 01.03.2019
• Subjects: association; support vector machine ( SVM); missing labels; multi-label learning
• ISSN: 1006-6748
• DOI: 10.3772/j.issn.1006-6748.2019.01.013

Multi-label learning is an active research area which plays an important role in machine learn-ing. Traditional learning algorithms, however, have to depend on samples with complete labels. The existing learning algorithms with missing labels do not consider the relevance of labels, resulting in label estimation errors of new samples. A new multi-label learning algorithm with support vector ma-chine ( SVM) based association ( SVMA) is proposed to estimate missing labels by constructing the association between different labels. SVMA will establish a mapping function to minimize the number of samples in the margin while ensuring the margin large enough as well as minimizing the misclassi-fication probability. To evaluate the performance of SVMA in the condition of missing labels, four typical data sets are adopted with the integrity of the labels being handled manually. Simulation re-sults show the superiority of SVMA in dealing with the samples with missing labels compared with other models in image classification.