(積ん読)Bias in error estimation when using cross-validation for model selection
Conclusion (section)
In our work, we observed this bias for two different resampling methods (10-fold CV and LOOCV).
This is effectively two nested CV loops; the outer loop estimates the generalization error while the inner CV is used for tuning the parameters.
「2つのネストしたCVループ」
「外側のループは汎化誤差を見積もり、内側のCVはパラメタチューニングに使われる」