Data calibration method and system based on manifold migration learning

Abstract

The invention relates to a data calibration method and system based on manifold migration learning. The data calibration method comprises the steps of obtaining feature data of a calibrated label to serve as a source domain, obtaining feature data of a label to be calibrated to serve as a target domain, performing principal component analysis on the source domain and the target domain to obtain asource feature vector and a target feature vector; respectively mapping the source feature vector and the target feature vector to a manifold space so as to obtain a source manifold feature of the source domain in the manifold space and a target manifold feature of the target domain in the manifold space; and counting the label types of the source domain, obtaining an average value of the manifoldfeatures under each label type according to the amount of the feature data under the label types, and performing label calibration for the feature data in the target domain according to the distancebetween the average value and the target manifold feature. The data calibration method simplifies the calibration of large-scale data, improves the generalization ability of the method and improves the operation efficiency of migration calibration.

Description

technical field [0001] The invention relates to the fields of machine learning, transfer learning and data calibration, in particular to a data calibration method and system based on manifold transfer learning. Background technique [0002] The era of big data has produced a large amount of user data in various aspects such as crowd behavior, traffic patterns, life data, health, office, and medical care. Based on these large-scale image, text, audio and video data, researchers can conduct more extensive and in-depth analysis and applications. At the same time, the industry can also customize more personalized services for users based on these data. However, although these data can be easily obtained, they are often presented in an uncalibrated form. Without adequate labeling, it is difficult to make maximum use of this data. Moreover, usually only some side information of these data can be obtained (for example, different images often present different feature distributio...

AI Technical Summary

Problems solved by technology

However, the existing transfer learning methods often only focus on solving two problems: either performing subspace learning, mapping data to different subspaces for different data adaptation; or performing probability distribution adaptation, in a high-dimensional Minimize the distance between the existing calibration data and the target data to be solved in space

[0008] 3) Existing transfer learning: After the subspace transfer learning method learns the subspace, the data features still drift, that is, the features no longer obey the same data distribution, resulting in inaccurate results; the probability distribution adaptation method is only carried out in the original feature space , and the features in the original space are often distorted, that is, the features extracted by the usual feature extraction methods cannot completely represent the characteristics of the original data, and the results are not accurate enough

Images