基于距离尺度学习的新类识别方法

New Class Recognition Method Based on Distance Metric Learning

在在线分类任务中经常会出现新类别,导致数据分布发生显著变化,使得已有分类器不再适用.如何识别新类以使分类器能适应其出现已成为在线分类亟待解决的问题.本文提出基于距离尺度学习的识别偏离型新类的算法用于解决该问题.该方法能在缺少先验知识的前提下自动识别新类,并较好地解决了样本间类别相似性同样本间距离不一致的问题,为分类器的自适应更新提供了关键技术.在多个数据集上的实验结果表明在客观新类出现后该方法能有效发现新类,可使更新后的分类器保持较高准确度,为实现适应新类的在线分类系统奠定坚实基础.

In online classification tasks, new class of patterns sometimes emerges, which makes the distribution change significantly and current classification models invalid. A method based on distance metric learning is proposed to recognize new class from existing classes without the apriori knowledge about emerging class. And it can make the class similarity represented by using distance between two objects, which is the key of promoting the performance of recognition. Therefore, the proposed method can be applied to the adaptive classification. The experimental results show that the proposed method can recognize new class well, and on the basis of this method the online classifier is adapted and it can predict the instance better than the original one.