CV score OR test score

以下是来自对于kaggle 大佬owen zhang的采访:

What is your approach to solid CV/final submission selection and LB fit?

This really depends on if there is a public LB data leak.
For time-sensitive models, it is very common that the public LB gives us “hints” about the private LB. In that case I use almost exclusively the public LB results as validation. If that is the case, I would usually give public LB more weight.
In the case the data is small/noise, I use a combination of CV and public LB.
I also usually choose 2 different models as final results -- one favors public LB and one favors CV.

以下来自于kaggle大佬radder的PPT:

这个问题要具体分析,实践中多数情况是最后一种,但是根据label定义以及样本的构造,要具体分析

留言

熱門文章