采用分而治之策略的快速多标签支持向量机分类算法研究  

Research on Fast Multi-Label SVM Classification Algorithm Using Divide-and-Conquer Strategy

作　　者：刘竞 郭忠文[1] 孙中卫 刘石勇 王续澎 LIU Jing;GUO Zhong-Wen;SUN Zhong-Wei;LIU Shi-Yong;WANG Xu-Peng(College of Information Science and Engineering, Ocean University of China, Qingdao 266100, China)

机构地区：[1]中国海洋大学信息科学与工程学院,山东青岛266100

出　　处：《中国海洋大学学报（自然科学版）》2020年第12期160-166,共7页Periodical of Ocean University of China

基　　金：国家自然科学基金项目(61379127)资助。

摘　　要：在大规模多标签分类中,繁重的计算复杂度已严重限制了非线性核支持向量机的使用。因此,本文提出了采用分而治之策略的快速多标签支持向量机分类算法。首先,采用二元关联问题转换策略将多标签分类问题转换为多个二元分类问题。然后,每个二元分类问题都可以被改进的采用分而治之策略的支持向量机分类算法解决,其改进体现在采用DEC(Different Error Cost)方法来克服标签数据不平衡问题。最后,通过集成每个二元分类问题解决方案来实现快速多标签分类。它在训练和测试速度、测试性能等方面优于其它快速多标签分类算法。在两组大规模多标签数据集上的实验结果表明:该算法的训练和测试速度是最快的,测试性能接近ML-LIBSVM分类算法,优于其它快速多标签分类算法。In large-scale multi-label classification,excessive computational complexity has severely restricted the application of non-linear kernel support vector machines(SVMs).Therefore,we proposed fast multi-label SVM classification algorithm using divide-and-conquer strategy(MLDC-SVM).which firstly uses the binary relevance problem transformation strategy to transform the multi-label classification problem into multiple binary classification problems.After that,each binary classification problem can be solved by using the improved binary SVM classification algorithm using divide-and-conquer stra-tegy.The improvement is reflected in the use of DEC(Different Error Cost)method to overcome the label data imbalance issue.Finally,it achieves fast multi-label classification by integrating the solutions of each binary classification problems.This algorithm is superior to other fast multi-label classification algorithms in training speed,testing speed and testing performance.Experimental results on two large-scale multi-label data sets show that the training and testing speed of MLDC-SVM classification algorithm is the fastest.Meanwhile,the testing performance of MLDC-SVM classification algorithm is similar to that of ML-LIBSVM classification algorithm,which is better than that of other fast multi-label classification algorithms.

关 键 词：多标签分类 支持向量机 非线性核 分而治之策略 标签数据不平衡 不同错误成本方法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]