一种改进的多源域多视角学习算法  

An Improved Multi-source and Multi-view Learning Algorithm

作　　者：刘桂峰[1] 赵志刚[1] 王福驰 吕慧显[2] 李建波[1] 李琳[1] 

机构地区：[1]青岛大学信息工程学院,青岛266071 [2]青岛大学自动化工程学院,青岛266071

出　　处：《青岛大学学报（自然科学版）》2015年第4期38-43,共6页Journal of Qingdao University(Natural Science Edition)

基　　金：山东省科学技术发展计划项目(批准号:2012YD01058)资助

摘　　要：在多源域迁移学习和多视角迁移学习的基础上,对基于多源域多视角的迁移学习算法进行了改进。该算法对分类错误率高的分类器不是仅仅强制设置成最大阈值,而是进行取反以获得更高的利用价值。在源域更新过程中,考虑TrAdaBoost的源域权重下降过快问题,加入动态因子,改变了源域权重收敛过快导致的权重熵过分转移到目标域的问题。实验结果表明,该算法在识别方面具有一定优势。Based on existing learning algorithms,a new improved multi-source and multi-view learning algorithm was proposed.The algorithms does not simply intend to set a maximum threshold value of the classifier whose classification error rate is high,but to invert it to acquire larger utilization value.With the problem that the source weights of TrAdaboost dropping too fast being considered dynamic factors were added to improve that problem which resulted in weight entropy drift to the target domain during the update of the source domain.Experimental results showed that the algorithm can perform well in recognition.

关 键 词：多源域 迁移学习 多视角 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]