从组合的randomForest回归对象计算R平方（解释的％Var）

Question

When calculating a randomForest regression, the object includes the R-squared as " % Var explained: ... ".

library(randomForest)
library(doSNOW)
library(foreach)
library(ggplot2)

dat <- data.frame(ggplot2::diamonds[1:1000,1:7])
rf <- randomForest(formula = carat ~ ., data = dat, ntree = 500)
rf
# Call:
#   randomForest(formula = carat ~ ., data = dat, ntree = 500) 
#                Type of random forest: regression
#                      Number of trees: 500
# No. of variables tried at each split: 2
# 
# Mean of squared residuals: 0.001820046
# % Var explained: 95.22

However, when using a foreach loop to calculate and combine multiple randomForest objects, the R-squared values are not available, as it is noted in ?combine :

The confusion , err.rate , mse and rsq components (as well as the corresponding components in the test compnent, if exist) of the combined object will be NULL

cl <- makeCluster(8)
registerDoSNOW(cl)

rfPar <- foreach(ntree=rep(63,8), 
                 .combine = combine, 
                 .multicombine = T, 
                 .packages = "randomForest") %dopar% 
                 {
                   randomForest(formula = carat ~ ., data = dat, ntree = ntree)
                 }
stopCluster(cl)

rfPar
# Call:
#   randomForest(formula = carat ~ ., data = dat, ntree = ntree) 
#                Type of random forest: regression
#                      Number of trees: 504
# No. of variables tried at each split: 2

Since it was not really answered in this question : Is it at all possible to calculate the R-squared (% Var explained) and Mean of squared residuals from an randomForest object afterwards?

(Critics of this parallelization might argue to use caret::train(... method = "parRF") , or others. However, this turns out to take forever. In fact, this might be useful for anybody who uses combine to merge randomForest objects...)

Answer 1

Yes. You can calculate the R-squared value after the fact by taking the predictions that result from your training data and your trained model and comparing them to the actual values:

# taking the object from the question:
actual <- dat$carat
predicted <- unname(predict(rfPar, dat))

R2 <- 1 - (sum((actual-predicted)^2)/sum((actual-mean(actual))^2))

Or Mean Squared Error:

caret::RMSE(predicted,actual)