基于域相关性与流形约束的多源域迁移学习分类算法

发布时间：2018-05-05 01:14

本文选题：迁移学习 + 多源域迁移　；参考：《计算机应用研究》2017年02期

【摘要】：针对传统机器学习要求训练样本和测试样本具有相同分布的假设在实际应用中难以满足的问题,进行了迁移学习的研究。当源域与目标域的相关性较小时,粗暴式地强制性迁移不仅不能促进目标域学习,甚至会降低学习性能,即出现负迁移问题。为此,提出了一种基于多重相似性的多源域迁移学习方法。该方法首先从域—域和样本—域两个层面来更加准确地衡量目标域与多个源域的相关性;然后根据目标域分类器与相关源域分类器对目标域无标记样本具有相似决策值的平滑性流形假设,实现源域知识向目标域迁移的目的。在公共数据集上的实验结果表明,所提出的迁移学习方法不仅能够从多个不同源域中挖掘更多的知识用于目标域学习,而且能够根据域间相似性有选择地进行迁移,可以有效地避免负迁移问题。
[Abstract]:In order to solve the problem that the traditional machine learning requires the same distribution of training samples and test samples, the transfer learning is studied. When the correlation between the source domain and the target domain is small, the rough mandatory transfer not only can not promote the learning of the target domain, but will even reduce the learning performance, that is, the problem of negative transfer occurs. Therefore, a multi-source domain transfer learning method based on multiple similarity is proposed. The method firstly measures the correlation between target domain and multiple source domains more accurately from two aspects: domain domain and sample domain. Then, according to the smooth manifold hypothesis that the target domain classifier and the correlation source domain classifier have similar decision value to the target domain unlabeled sample, the goal of knowledge migration from the source domain to the target domain is realized. Experimental results on common data sets show that the proposed transfer learning method can not only mine more knowledge from multiple different domains for target domain learning, but also selectively migrate according to the similarity between domains. The problem of negative migration can be avoided effectively.
【作者单位】：电子工程学院;安徽省电子制约技术重点实验室;
【基金】：国家“863”计划资助项目安徽省自然科学基金资助项目(1308085QF99,1408085MKL46)
【分类号】：TP181

【相似文献】