The Research On Online Transfer Based On Cost-Sensitive And Kernel For Classification

A large amount of online data is generated in our practical applications such as anomaly detection and credit card fraud detection.The lack of tag information in these online data and the information hidden in the online data changing of time pose great challenges for online data analysis tasks.In recent years,online transfer learning,which use labeled source domain information to help the learning of target online data has received extensive attention.However,the existing online transfer learning methods mostly update the model according to the classification accuracy,which is difficult to achieve the desired effect in practical applications.Therefore,this paper conducts online transfer learning methods for different application areas and the main research works are as follows:1)The definition of online transfer learning is introduced,and the current mainstream online transfer learning methods are analyzed and introduced in this paper.2)In online applications such as credit card fraud,the classification cost is more important than the accuracy rate,and the cost of online data usually changes over time.To solve above problems,this paper proposes an online transfer learning method with adaptive cost(OLAC)by taking the classification cost and accuracy into account.Firstly,a label distribution parameter is introduced to classification model and used for adaptive calculating the cost.Secondly,the source domain and the target domain are combined through the weighting parameters to realize the online transfer from the source domain model to the target domain.Finally,a classifier model is constructed according to taking the cost and accuracy into consideration simultaneously.Experimental results show the proposed approach can achieve better classification accuracy and minimum cost than several state-of-the-art baseline methods.3)In consideration of the fact of most linear classification models cannot solve nonlinearly separable online data and the kernel method in online environment has the problem of support vector boundaryless growth,we propose online transfer learning algorithm based on kernel(KOTL)for classification.The method directly maps the objective function in reproducing kernel Hilbert space to ignore whether the original target domain data is linearly separable in low-dimensional space,and use a buffering strategy to solve the support vector boundaryless growth problem of the kernel function.Finally,a classifier is updated by simultaneously optimizing the structural risk functional,the distribution matching between domains and the manifold constraint.Extensive experiments demonstrate that KOTL can obtain excellent performance on several image datasets.