非平衡样本分类的集成迁移学习算法

摘要

According to the auxiliary training data with large redundancy and imbalance between positive and negative samples, an improved integrated transfer learning algorithmic -The Unbalanced Integrated Transfer Learning Algorithmic is proposed. Applied these auxiliary training data to transfer and help classifying on target data.New sample initialization and regulation weight method highlighted negative sample identification ability. Through dynamic adjusting auxiliary training set, eliminated redundant data according to the weight lower threshold,reduced their influence on the classifier and improved the transfer learning's performance. Experimental results on the actual bridge monitoring data show that this algorithmic is advanced than TrAdaboost.%针对冗余数据量大且正负样本不平衡的辅助训练数据,提出了一种改进集成迁移学习算法,利用这些辅助训练数据迁移帮助目标数据进行分类.新的样本初始权重分配及调整策略,突出了对负样本的识别能力.通过动态调整辅助训练集,根据设定好的权重阈值下限适时地淘汰冗余数据,降低了冗余数据对分类器性能的影响,提升了迁移学习对非平衡样本的学习能力.本文利用桥梁实际监测数据进行的实验表明了该算法较TrAdaboost算法的有效性.