Method And Application Of Multi-source Transfer Learning Based On Deep Neural Network

Because of allowing the distribution varies between training and testing data,transfer learning is receiving more and more attention.Domain adaptation and domain generalization are two representative technologies of transfer learning.Multi-source transfer learning explores how to extract knowledge from multiple source domains and apply it to the target domain in a balanced way.It is the key technology to improve the performance of the model and promote the implementation of transfer learning.Although many methods and technologies have emerged in the field of multi-source transfer learning,there are still some problems to be solved due to the complexity of the multi-source problem.For example,the model needs to ensure the features have task-relevant discriminative information while extracting the universal features,the different data distribution differences between each source domain and the target domain should be taken into account,and the scheme to expand the application fields of multi-source transfer learning.Based on the above background,this thesis focuses on the method design and application expansion of multi-source transfer learning technology.By analyzing and summarizing the existing research of multisource transfer learning,and mainly focusing on the more complex problem of domain generalization,a new multi-source transfer learning network structure is proposed,and it has been applied on the practical economic management issue.The contributions of this dissertation are summarized as follows:(1)This thesis systematically studies and summarizes the relevant technology and methods of multi-source transfer learning,and forms a general framework of the network structure of deep multi-source transfer.This thesis also analyzes the current mainstream multi-source transfer learning methods.(2)This thesis proposes a domain generalization method based on source-specific adversarial learning.Multiple classifiers,multiple domain discriminators and a distribution generator are introduced,and a reference distribution is adaptively generated in the training process.At the same time,the domain-invariant features between different source domains and reference distribution are extracted respectively through domain-specific adversarial learning,and different classifiers are trained using this feature.The outputs of classifiers are aligned and weighted.Comparative experiments,ablation experiments and expansion experiments are conducted on the classical image datasets of domain generalization.The method can learn effective feature representation,and achieves remarkable generalization performance.(3)Under the actual application scenario of personal loan default prediction,this thesis introduces a method for identifying and dividing latent multi-source domains based on clustering algorithm and domain discrimination features.The method is integrated into the existing domain generalization network structure to realize the application of domain generalization in economic management issues.On the existing mixed domain datasets,the domain generalization method which integrates mixed latent domain division has achieved significant improvement without using domain labels.