论文部分内容阅读
针对源领域数据和目标领域数据分布类似的情况,提出一种基于多源动态TrAdaBoost的实例迁移学习方法.考虑多个源领域知识,使得目标任务的学习可以充分利用所有源领域信息,每次训练候选分类器时,所有源领域样本都参与学习,可以获得有利于目标任务学习的有用信息,从而避免负迁移的产生.理论分析验证了所提算法较单源迁移的优势,以及加入动态因子改善了源权重收敛导致的权重熵由源样本转移到目标样本的问题.研究结果表明,所提算法在2个和3个源领域的迁移学习精确度最高值分别达到90.7%和92.2%,能够得到较高的分类精度,更适于实例迁移学习.
Aiming at the similar situation of data distribution in source area and target area, a case-based migration learning method based on multi-source dynamic TrAdaBoost is proposed.Considering knowledge of multiple source fields, the learning of target tasks can make full use of all source domain information, and each training Candidate classifier, all the samples in the source domain are involved in learning, which can be used to obtain useful information for the target task learning, so as to avoid the negative migration.The theoretical analysis proves the advantages of the proposed algorithm over the single source migration and the improvement of the dynamic factor The weight entropy caused by convergence of source weight is transferred from the source sample to the target sample.The research results show that the maximum precision of the proposed algorithm in two and three source fields reaches 90.7% and 92.2%, respectively, Higher classification accuracy, more suitable for instance migration learning.