基于边际Fisher准则和迁移学习的小样本集分类器设计算法

如何利用大量已有的同构标记数据（源域）设计小样本训练数据（目标域）的分类器是一个具有很强应用意义的研究问题. 由于不同域的数据特征分布有差异，直接使用源域数据对目标域样本进行分类的效果并不理想. 针对上述问题，本文提出了一种基于迁移学习的分类器设计算法. 首先，本文利用内积度量的边际Fisher准则对源域进行特征映射，提高源域中类内紧凑性和类间区分性. 其次，为了筛选合理的训练样本对，本文提出一种去除边界奇异点的算法来选择源域密集区域样本点，与目标域中的标记样本点组成训练样本对. 在核化空间上，本文学习了目标域特征到源域特征的非线性转换，将目标域映射到源域. 最后，利用邻近算法（k-nearest neighbor，kNN）分类器对映射后的目标域样本进行分类. 本文不仅改进了边际Fisher准则方法，并且将基于自适应样本对 筛选的迁移学习应用到小样本数据的分类器设计中，提高域间适应性. 在通用数据集上的实验结果表明，本文提出的方法能够有效提高小样本训练域的分类器性能.

Classifier-designing Algorithm on a Small Dataset Based on Margin Fisher Criterion and Transfer Learning

It has great practical significance to design a classifier on a small dataset (target domain) with the help of a large dataset (source domain). Since feature distribution varies on different datasets, the classifiers trained on the source domain cannot perform well on a target domain. To solve the problem, we propose a novel classifier-designing algorithm based on transfer learning theory. Firstly, to improve the compass of the same category and separateness of different categories in the source domain, this paper utilizes the extended margin Fisher criterion where the distance is measured by the inner product between data. Secondly, to select good sample pairs for transfer learning, this paper presents an algorithm to get rid of marginal singular points by selecting high-density samples in the source domain. The non-linear feature transformation mapping the target domain to the source domain is learned in the kernel space. Finally, k-nearest neighbor (kNN) classifiers are trained for classification. Compared with the existing works, this paper not only extends the margin Fisher criterion, but also applies the transfer learning theory based on the algorithm of selecting training sample pairs to design classifiers of a small dataset. We experimentally demonstrate the superiority of our method to effectively improve the performance of classifiers on the general datasets.