A Systematic and Generic Correlation-Based Design Approach for Data Sample Reduction in ML-Training
Xin-Yu Shih, Ming-Jyun Wu, Hsiang-En Wu · 2022 IEEE International Conference on Consumer Electronics - Taiwan · 2022
In this paper, we propose a generic design approach for efficiently reducing data samples. The significant design spirit is to apply and check the correlation relationship between each feature and the classified label. It is applicable for any types of machine-learning models, providing 5-tuple parameters. By adjusting the design parameters, our systematic approach can achieve the trade-off between classification accuracy and sample reduction ratio while satisfying the users' demand. In the verification with suggested parameter setting and different ML models, the data samples are reduced by 15%–25% under the acceptable accuracy loss.