为什么h2o包中的交叉验证数据存在不同的MSE

这两个MSE值的计算方式不同。

第一个(0.1641124)是在交叉验证期间使用保持集上的所有预测计算得出的:

创建模型:

m <- h2o.glm(x = 2:5,
             y = 1,
             train,
             nfolds = 10,
             seed = 123,
             keep_cross_validation_predictions = TRUE,
             keep_cross_validation_fold_assignment = TRUE)

preds <- as.data.frame(h2o.cross_validation_holdout_predictions(m))

计算MSE:

mean((preds$predict - as.data.frame(train)$Sepal.Length)^2)
#output
0.1641125

其中,较低的MSE(0.14977892)表示每个保持组的MSE平均值:

folds <- as.data.frame(h2o.cross_validation_fold_assignment(m))

library(tidyverse)
data.frame(preds = preds$predict,  #create a data frame with hold out predictions
           folds = folds$fold_assignment,  #folds assignement
           true = as.data.frame(train)$Sepal.Length) %>% #true values
  group_by(folds) %>% #group by folds 
  summarise(mse = mean((preds - true)^2)) %>% # calculate mse for each fold
  ungroup() %>%
  summarise(mse = mean(mse)) %>% #average them
  as.numeric
#output
0.1497789

library(h2o)

h <- h2o.init()
data <- as.h2o(iris)
part <- h2o.splitFrame(data, 0.7, seed = 123)
train <- part[[1]]
test <- part[[2]]