基于SVM期望间隔的多标签分类的主动学习

Active Learning for Multi-label Classification Based on SVM's Expect Margin

分类是数据挖掘领域研究中的核心技术之一。得到一个性能良好的分类器需要大量的训练样本,而对样本进行标记是一个十分消耗资源的过程,对多标签样本进行标记就更加困难。为了尽可能降低标记样本的成本,需要找出最能代表类别信息的样本。在基于SVM的分类方法中,分类器间隔越大,分类的精度就会越差。提出了一种基于期望间隔的主动学习方法,即依据当前分类器,选择最快缩小分类间隔的样本。通过实验证明,基于期望间隔的学习策略比基于决策值以及基于后验概率的策略有着更好的学习效果。

Classification is one of the key techniques of data mining.It requires a large number of training samples to obtain a favorable classifier,but it is resource-consuming to create label for each sample,it is even more so for multi-label samples.In order to reduce costs,it should find the most informative samples which can represent the classes.The classifiers which are based on SVM,the larger margin,the classifier＇s accuracy will be poorer.This paper proposed an active learning method based on SVM＇s expect margin which relies on current classifier,select samples that can reduce classifier＇s margin fastest.The experimental results show that the method based on expect margin outperforms than other active learning strategy based on decision value and posterior probability strategy.