Data set equalization training method based on artificial intelligence

By generating an initial class feature matrix, estimating the minority class density, and performing local dimension analysis of the manifold, combined with conditional generative adversarial networks and manifold-preserving distance metrics, the imbalanced data problem was solved, and the model's minority class identification accuracy and generalization ability were improved.

CN120974186APending Publication Date: 2025-11-18BEIJING QINGYUAN CHUANGYAN TECHNOLOGY CO LTD
0 Cites 0 Cited by

Patent Information

Application Number
CN202511081792.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-08-04
Publication Date
2025-11-18

AI Technical Summary

Technical Problem

Existing technologies fail to effectively consider the manifold structure features and intrinsic dimensionality distribution characteristics of minority class samples in the feature space when dealing with the problem of imbalanced data. This leads to a mismatch between synthetic samples and real samples, affecting the model's recognition performance and generalization performance.

Method used

An initial class feature matrix is ​​generated through feature extraction and class classification. Minority class density estimation and manifold local dimension analysis are performed. Candidate samples are generated using a conditional generative adversarial network. Deviation samples are eliminated by manifold-preserving distance metric and deviation filtering mask. Finally, class reweighting training is performed to form a balanced dataset.

Benefits of technology

It improves the accuracy of minority class identification and the model's generalization ability in imbalanced data scenarios, enhances the model's robustness, and ensures the consistency between synthetic samples and real samples.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention discloses a data set equalization training method based on artificial intelligence, and particularly relates to the technical field of artificial intelligence. According to the method, an initial category feature matrix is formed by collecting a target data set and performing feature extraction and category classification; generating a feature collapse risk index and a low-dimensional manifold description through minority class density estimation and manifold local dimension analysis; generating a candidate sample set in a minority-class manifold neighborhood according to the feature collapse risk index; the deviation degree of the candidate samples is calculated through manifold maintaining distance measurement, the candidate samples with the deviation degree exceeding a preset deviation degree threshold value are removed, and a synthetic sample set is obtained; a synthetic sample set and a target data set are fused, and equalization training is performed based on a category reweighting strategy, so that feature space collapse and inter-class confusion are significantly reduced, and the recognition precision and generalization performance of the model under unbalanced data are improved.
Need to check novelty before this filing date? Find Prior Art