摘要： 现有多标签分类问题普遍被转换成多类分类问题，计算量较大，运行时间较长，且面对新类别加入时，拓展性较差。为此，提出一种基于球结构支持向量机的多标签分类方法。每一类别标签对应一个球域结构，提取球重叠区域的样本，依据距离差值度量样本类别相似度，确定样本所属类别。实验结果表明，该方法可以节省210 ms的训练时间，使平均查全率提高3.2%，适合大量样本分类。

关键词: 支持向量机, 距离差, 多标签分类, 多类分类, 主动学习, K折交叉验证

Abstract: Multi-label classification problem is generally converted into multi-class classification problem, and it can increase the amount of computation and extend running time. When a new class is added, the method has a poor expandability too. To solve this problem, this paper proposes a multi-label classification method based on sphere structured Support Vector Machine(SVM). Each category labels correspond a ball domain structure, this paper extracts the samples of ball overlap region, according to the distance difference measures the sample category similarity, and is to determine the category of samples. Experimental results show that the method can save 210 ms training time, improves the average recall rate of 3.2% and is suitable for a large number of sample classification.

Key words: Support Vector Machine(SVM), distance difference, multi-label classification, multi-class classification, active learning, K-fold cross validation

中图分类号: 

• TP309