tanh As a robust feature scaling method in training deep learning models with imbalanced data

Aijia Yang, Min Zhang, Huai Chen, Taihao Li, Shupeng Liu, Xiaoyin Xu

Abstract: Highlights•Designed a feature scaling technique to train learning models on imbalanced data.•The technique maps data to the linear region of tanh curve and suppresses outliers.•The scaled data are bounded between −1 and 1 and have bounded variations.•Experiments show that the technique surpasses other methods in deep learning.•Tests indicate that the technique has a robust capability in avoiding gradient explosion.

External IDs:doi:10.1016/j.patcog.2025.111746