UnbalanceClassify

Unbalance Classify

Class-Balanced Loss Based on Effective Number of Samples [CVPR 2019] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, Serge Belongie.
- a novel sample method using in loss calculate to solve unbalance problem.
- Two proposition about effective number of sample
  - $E_{n}=\left(1-\beta^{n}\right) /(1-\beta)$, where $\beta=(N-1)/N$
  - $$E_{n}=\left(1-\beta^{n}\right) /(1-\beta)=\sum_{j=1}^{n} \beta^{j-1}$$
  - N more, $\beta$ -> 1; N=1, $\beta$ -> 0
- use in softMax cross entropy, sigmoid, focal loss
Dice Loss for Data-imbalanced NLP Tasks [-] Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, Jiwei Li.
- To solve imbalance issue in NLP
- using Dice Loss
- like F1-score, attach similar FP == FN