R：使用cv.glm计算弹性网预测误差

Question

library(glmnet)
library(boot)
data(iris)
x <- model.matrix(Sepal.Length~., iris)[,-1]
y <- iris$Sepal.Length
m <- cv.glmnet(x, y)
> cv.glm(x, m, K = 10)
Error in UseMethod("predict") : 
  no applicable method for 'predict' applied to an object of class "c('matrix', 'double', 'numeric')"

or

bestLambda = m$lambda.min
m2 <- glmnet(x, y, family = "gaussian", lambda = bestLambda)
>cv.glm(x, m2, K = 10)


 Error in glmnet(x = x, y = y, family = "gaussian", lambda = bestLambda,  : 
  unused argument (data = c(3.5, 3, 3.2, 3.1, 3.6,

In reference to this question, I'm trying to obtain the K-fold cross-validated prediction error of my Elastic net model using cv.glm , however, I can't seem to do so due to the error. I'm not quite sure if the cv.glm function can be used to calculate the prediction error of a class cv.glm or glmnet object.

Answer 1

I think you are mixing up glm with glmnet (elastic net, with lasso & ridge penalties). cv.glm expects a glm model, not a glmnet model.

Try either of the following:

1. Use glmnet to compute k-fold cross-validation errors with cv.glmnet like the following:

    library(glmnet) library(boot) data(iris) x <- model.matrix(Sepal.Length~., iris)[,-1] y <- iris$Sepal.Length m <- cv.glmnet(x, y, nfolds=10) m$lambda.min #[1] 0.0003839539 m$lambda.1se #[1] 0.009078549 plot(m$lambda, m$cvm,type='l', xlab=expression(lambda), ylab='CV errors', main=expression(paste('CV error for different ', lambda))) lines(m$lambda, m$cvup, col='red') lines(m$lambda, m$cvlo, col='red') 

   

[EDITED]

prediction error on the training dataset:

mean((y-predict(m, newx=x))^2)
# [1] 0.1037433

2. Fit a glm model and use cv.glm to compute the cross-validation error delta (without regularization). As per the documentation of cv.glm :

delta: A vector of length two. The first component is the raw cross-validation estimate of prediction error. The second component is the adjusted cross-validation estimate. The adjustment is designed to compensate for the bias introduced by not using leave-one-out cross-validation.

df <- cbind.data.frame(x, y)
m <- glm(y~., df, family='gaussian')
cv.glm(df, m, K = 10)$delta 
# [1] 0.09992177 0.09940190