冗余数据裁剪的自适应迁移学习算法

Adaptive transfer learning algorithm by redundant data clipping

为提高Tradaboost迁移学习的适应性,研究目标领域和源领域的样本权值。考虑目标域和源域间的样本权值差距偏大、负迁移明显的缺陷,提出一种RCTRA算法。增加源领域样本权值回补参数,利用动态冗余数据裁剪算法,小于初始权值的源域样本不可选,在可选数据中动态剔除权值小于设定阈值的数据。通过测试青荣城际高速铁路、青威客车客流数据进行验证,实验结果表明,RCTRA算法较传统迁移学习算法具有更好的有效性和鲁棒性。

To improve the adaptability of Tradaboost algorithm,the weight of the samples in the target domain and the source domain was studied.A RCTRA algorithm was proposed.The source domain sample weighting compensation parameter was reduced for the defects that the sample weight is large between the target domain and the source domain and the negative transfer is obvious.Using the dynamic redundant data clipping method,the source domain samples less than the initial weight were not available and the data were further removed with the weight less than the set threshold value in the optional data.The data of Qing Wei passenger bus and high-speed railway of the Qing Rong city were tested.Experimental results show that the RCTRA algorithm is more effective and robust than the traditional transfer learning algorithm.