A Solution to Address the Issue of Model Misjudgment and Omission Caused by AI Training Data Bias in Depth
Journal of Electrical Systems · 2026-06-05 · 22 (01) · 584-589
DOI: 10.52783/jes.9653 ↗At a glance
Journal article in Journal of Electrical Systems, published online on 2026-06-05. Written by Deze Wang. The full text is freely available (open access, CC BY-ND 4.0).
Abstract (original language)
Addressing false positives and false negatives caused by biased AI training data requires a comprehensive solution. At the data level, broadening data sources ensures diversity; weighted sampling, oversampling, and synthetic data are used to enhance balanced distribution; biased samples are eliminated, and labeling standards are standardized. During training, a fairness loss function and adversarial training are introduced. Using a fixed random seed ensures stability, and domain knowledge graphs are used to supplement context. In the iteration phase, a false positive case knowledge base is built, and samples are continuously collected through log analysis and user feedback. Root cause analysis is used to guide model fine-tuning. Simultaneously, a dynamic evaluation and closed-loop iteration mechanism is established to monitor fairness metrics and ensure the model’s generalization ability, comprehensively reducing the impact of bias from source to application.
Cite
APA
Wang, D. (2026). A Solution to Address the Issue of Model Misjudgment and Omission Caused by AI Training Data Bias in Depth. Journal of Electrical Systems, 22(01), 584-589. https://doi.org/10.52783/jes.9653
Vancouver
Wang D. A Solution to Address the Issue of Model Misjudgment and Omission Caused by AI Training Data Bias in Depth. Journal of Electrical Systems. 2026;22(01):584-589. doi:10.52783/jes.9653
BibTeX
@article{Wang2026,
author = {Wang, Deze},
title = {A Solution to Address the Issue of Model Misjudgment and Omission Caused by AI Training Data Bias in Depth},
journal = {Journal of Electrical Systems},
year = {2026},
volume = {22},
number = {01},
pages = {584--589},
doi = {10.52783/jes.9653}
}
EUVIMED