Research On Domain Adaptation Method Based On Feature Selection

Domain adaptation,as the main research branch of transfer learning,can help the classifi-cation model extract knowledge from the data in the source domain and predict the target task in the target domain with different distribution.With the development of Internet technology,rich data sources have brought a huge amount of data,which has helped the traditional machine learning models and caused obstacles.The main obstacle is that the diversification of data caus-es the distribution of data fields to differ,which affects the effect of machine learning models.The research of domain adaptation methods is to solve or reduce such obstacles universally.In the past domain adaptation methods,researchers often learn one or more feature mapping functions by means of feature dimension reduction or subspace learning to achieve the goal of small differences in domain distribution and classification of features.In order to be able to take advantage of the feature selection methods in previous studies,we use feature selection as a means and use in the domain adaptation research.The main work of this paper includes the following:1)This paper has made a detailed analysis of the theoretical background and the correlative method of domain adaptation,and clarified the nature of domain adaptation problem.On this basis,the key measurement methods are reviewed.At the same time,the related work of the feature selection research is sorted out.2)The TFS method is proposed.By minimizing an objective function including domain distribution difference metric,training error loss function and l₀-norm constraint,at last the transferable and discriminative features is obtained.3)Considering the adaptation of conditional distribution based on only the maginal distri-bution adaptive,and adding a learnable balance factor between the two distribution metrics in order to improve the robustness of the model target,an BTFS method with adaptive balanced distribution is obtained.4)A distributed optimization framework is proposed,which can effectively solve the pro-posed model,that is,solve a mixed integer programming problem with l₀-norm constraints.5)The effects of the proposed method are demonstrated on a variety of real datasets by comparison with benchmark methods,visualization of feature subsets and number of iterations.From the perspectives of negative transfer and target effectiveness,the robustness and effective-ness of domain adaptation methods based on feature selection are demonstrated.