| Transfer learning is a method of studying how to use the existing experience and knowledge in other related fields to help learn the target task.Most of the existing researches on methods are conducted on offline data.However,practical applications often need to face online tasks.The training samples in these problems come in a certain order,which are usually not directly available or require a high price.Therefore,research on effective online transfer learning algorithms has important practical significance.However,the existing online transfer learning methods simply implement the introduction of online learning methods into the problem of transfer learning,and do not fully consider whether the selected online learning strategy is suitable for the current transfer challenges and the new performance problems that result from it.Including the problem that the online transfer model cannot capture the more complex and richer association structure information in the sample,weight drift problem,Overfitting of misclassified data and imbalance of source and target data.Therefore,this article aims to combine ensemble learning and transfer learning to construct three new online transfer learning algorithm to solve the above-mentioned problems and improve the classification performance.The first algorithm we propose is the Dynamic Compensation strategy for Online Transfer Boosting algorithm,DCOTB.Boosting of ensemble learning is used to build a basic classifier,and a dynamic weight compensation strategy is innovatively introduced to solve the weight drift problem in online transfer ensemble learning.The second algorithm we propose is the Online Transfer Bagging algorithm,OTBag.The algorithm combines the online Bagging idea in ensemble learning,and uses bootstrap sampling to replace weight adjustments to avoid overfitting of misclassifications and weight drift problems.The algorithm also uses OTBag-SDMV and OTBag-JDSMV two filtering strategies to improve the algorithm’s ability to deal with negative transfer.Finally,we propose an Online Transfer algorithm based on MultiBoosting,called OTMB.The algorithm draws on the performance of Bagging and AdaBoost in reducing bias and variance,and retains this advantage through their combination,so as to effectively reduce the problem of overfitting to wrong data under a single Boost model,and the problem of sample loss and imbalance between source data and target data under a single bagging method.The paper verify the effectiveness of the three algorithms in turn through the experimental results on the data set. |
