18

To fit a lasso model using glmnet, you can simply do the following and glmnet will automatically calculate a reasonable range of lambda values appropriate for the data set:

glmnet(x, y, alpha = 1)

I know I can also do cross validation natively using glmnet. However, I would like to use the caret package so I can train and compare multiple models in a unified fashion. The problem is I don't know how to fix only the alpha parameter when calling glmnet from caret because caret wants to tune over both alpha and lambda.

train(y ~ ., data = train, method = 'glmnet', trControl = ctrl, tuneGrid = data.frame(alpha = 1))
Error in train.default(x, y, weights = w, ...) : 
  The tuning parameter grid should have columns alpha, lambda

Is there any way in general to specify only one parameter and allow the underlying algorithms to take care of the unspecified parameters according to the default methods of the algorithms?

EDIT:

The reason I would like to let glmnet to choose the lambda is that the appropriate range for lambda can vary a lot and glmnet does a good job of choosing a good range:

library(ElemStatLearn)
library(glmnet)

dat <- prostate

train   <- subset(dat,  train, select = -train)
train.x <- as.matrix(subset(train, select = -lpsa))
train.y <- train$lpsa

print(glmnet(train.x, train.y, alpha = 0)$lambda)
print(glmnet(train.x, train.y, alpha = 1)$lambda)

You can see that the ranges of lambda are very different. To use custom model in train, I thought about running glmnet first and then assign the calculated lambda to the grid: 

glmnetGrid <- function(x, y, len = NULL) {
  library(glmnet)
  lam <- glmnet(x, y, alpha)$lambda
  expand.grid(lambda = lam)
}

But then I need to pass alpha to the grid function and I don't think the framework allows that. Also it seems inefficient to run glmnet one more time just to get the lambda. Is there a way to simply let glmnet run and assign its lambda to the grid so caret can use the lambda values in subsequent analysis?

  • You can write your own method that only optimizes `alpha`. See [this page](http://caret.r-forge.r-project.org/custom_models.html) for details and examples. – topepo Mar 05 '14 at 00:20
  • 4
    @topepo: I actually don't want to only optimize `alpha`. I want to assign a fixed `alpha` to fit only ridge or only lasso, but have `caret` use the `lambda` that's generated from `glmnet` as the grid. The custom model approach doesn't seem like it will work. Please see my edit. Thanks. –  Mar 05 '14 at 03:23
  • 1
    I think your best bet might be to just not use caret and roll your own solution here. It looks like you're probably ultimately going to have a fairly complicated solution to accomplish something reasonably simple. Caret is great because it's convenient. In your case, it doesn't seem particularly convenient, so you probably don't want to use it. It would probably take you way less time to just build what you need yourself than to figure out how to get caret to bend to your will, assuming it's even capable of doing what you're trying. – David Marx Mar 05 '14 at 03:55
  • Another advantage of `caret` is that it allows you to use a custom model performance metric. `cv.glmnet` only has a few metrics available, which don't always suit the prediction task at hand. Here's a solution I worked out that I think matches what `glmnet` does:, where X is your predictor matrix, Y is your response, and alpha is your chosen value: `lamda.max – rw2 Jun 21 '21 at 08:33

0 Answers0