为什么h2o包中的交叉验证数据存在不同的MSE？

Question

在摘要输出中，交叉验证数据的MSE为0.1641124，而在详细的交叉验证度量摘要中为0.14977892。它们不是相同的度量标准吗？

library(h2o)

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

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


#...
#H2ORegressionMetrics: glm

#** Reported on cross-validation data. **
#** 10-fold cross-validation on training data (Metrics computed for combined 
#holdout predictions) **

#MSE:  ***0.1641124***
#RMSE:  0.4051079
#... 

#Cross-Validation Metrics Summary: 
#  mean  sd  cv_1_valid cv_2_valid cv_3_valid  cv_4_valid  cv_5_valid cv_6_valid  cv_7_valid cv_8_valid cv_9_valid


#...

#  mse  ***0.14977892*** 0.053578787  0.14102486 0.14244498 0.05266633  0.19028585 0.043878503 0.12635022  0.13820939 0.15831167 0.33359975

Answer 1

这两个MSE值计算不同。

第一个(0.1641124)是使用交叉验证期间对保持集的所有预测来计算的：

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)

提取出预测

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

较低的最小均方值(0.14977892)表示每个保持集的平均均方值：

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

要复制第一次运行：

library(h2o)

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