Confrontation migration learning method and system for cross-border big data analysis

Abstract

The present invention provides an adversarial transfer learning method and system for cross-border big data analysis. The method includes: substituting the random multi-linear fusion representation corresponding to each tensor in the tensor set corresponding to the unlabeled data set in the source domain and the target domain into the original loss function of the discriminator to obtain the current loss function of the discriminator, and Using backpropagation, adjust the parameters of the discriminator to minimize the current loss function as the current best loss function of the discriminator; the tensor in the tensor set is in the preset deep neural network The tensor product of the data vectors of all data layers in the preset data layer set; update the preset deep neural network parameters based on the current best loss function and enter the next update of the preset deep neural network parameters until the parameters converge . The joint distribution offset of multiple data layers in the preset deep neural network preset data layer set obtained by the present invention is reduced, and the application effect is better in the target field.

Description

technical field [0001] The present invention relates to the technical field of data analysis, and more specifically, to an adversarial migration learning method and system for cross-border big data analysis. Background technique [0002] Among many machine learning tasks, the deep neural network method is currently the best method. However, the deep neural network can only obtain good task results after supervised learning training after obtaining enough rich labeled data. In order to obtain a deep neural network with better effect for completing the target task when there is less labeled data in the target domain, cross-domain learning is usually used to use the abundant labeled data in the source domain for the target domain. Acquisition of Deep Neural Networks. Under the deep neural network based on the rich labeled data in the source domain, there is a problem of distribution deviation between the data in the source domain and the data in the target domain, so the effe...

AI Technical Summary

Problems solved by technology

[0005] The present invention provides an adversarial transfer learning method and system oriented to cross-border big data analysis, in order to overcome the fact that the data in the source domain and the data in the target domain may still be the same under several data layers at the top layer of the deep neural network obtained by the existing adversarial transfer learning method. There is a data distribution offset, and when the data distribution of the source domain and the target domain presents a multi-mode complex structure, it may be difficult for the deep neural network to capture the complex data distribution characteristics to align the distribution fine-grained, so that the source domain and the target domain. The data distribution shift is still large, and the effect of applying the target field to complete the target task is not good.

Images