传统主动学习中采用的批量采样模式忽略了样本之间的相互关系,因而会不可避免地引入冗余。针对上述问题,提出了一种动态批量采样模式,采取"逐一标注,批量训练"的流程,综合利用当前分类模型和先前标注样本对后续采样进行动态指导;在此基础上,进一步提出了基于动态确定度传播的选择性采样算法,有效地提高了所选取样本的信息量。实验结果证明,基于动态确定度传播的选择性采样算法能够显著改进分类结果。
In traditional active learning,selective sampling was performed in batch mode,which neglected examples' correlation and thus inevitably brought in redundancy.This paper presented a dynamic batch sampling mode,using both the existing classification boundary and the previously labeled examples as guidance for further selection.Then it proposed a dynamic certainty propagation(DCP)algorithm for informative example selection.Experimental results demonstrate the effectiveness of selective sampling with DCP algorithm.