基于多源动态TrAdaBoost的实例迁移学习方法

Instance-based transfer learning method using multi-source dynamic TrAdaBoost

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.

This study used the distribution of source data and target data to propose the instance-based transfer learning method on the basis of multi-sources dynamic TrAdaBoost. This paper took multi-domain knowledge into consideration to make the target task learning use all source information. All the source domain samples took part in learning in order to obtain useful information of beneficial target task learning, and avoided negative transfer when training candidate classifier. Theoretical analysis verifies that： the proposed algorithm is better than single-source transfer; and by means of adding the dynamic factor, this algorithm improves the defect that weight entropy to drift from source to target instances. The results show that transfer learning accuracy from two and three domains of the proposed method reaches 90.7 % and 92.2 %, respectively. The proposed method can obtain higher classification accuracy, which is much applicable for instance transfer learning.