数据分析：基于K-folds + repeats tuning 模型参数

前言

为了得到robust parameters，常常会对模型参数进行tuning，tuning的方法有K-folds CV或 N repeats 构建模型，前者对单次运行时划分样本训练模型，后者重复前者N次，接着选择前者最小lambda组合成最小lambda集合，再根据median或min等选择最佳lambda，最后基于最佳lambda构建新模型。

Codes

library(glmnet)
data(QuickStartExample)

df_lambdas_min <- c()
# 10-fold CV + 10 repeat
for(i in 1:10){
  cvfit <- cv.glmnet(x=x,
                     y=y,
                     nfolds = 10,
                     alpha = 1,
                     nlambda = 100,
                     type.measure = "auc")

# require(ipflasso)
# cvfit <- cvr.glmnet(X=dat_table,
#                  Y=dat_target,
#                  family='binomial',
#                  nfolds = 10,
#                  alpha = 1,
#                  ncv = 10,
#                  nlambda = 100,
#                  type.measure = "auc")

  df_lambdas_min <- rbind(df_lambdas_min, cvfit$lambda.min)
}
print(df_lambdas_min)

Notes: K folds设置应该考虑到sample size的问题，每个fold的sample size一定要大于8，所以10 folds的最小sample size是80

1. sample size per condition less than 8

   
   

2. sample size per folds less than 10

   
   

参考

1. Circulating Protein Biomarkers for Use in Pancreatic Ductal Adenocarcinoma Identification

2. An Introduction to glmnet

3. Repeating cv.glmnet

参考文章如引起任何侵权问题，可以与我联系，谢谢。