Error attempting Logistic Regression

[dhimmel]

I am encountering an error when trying to run a binomial logistic ridge regression. The error does not occur when using the default least squares regression.

import GLMNet
cv_guass = GLMNet.glmnetcv(X, y, GLMNet.Normal(), alpha=0.0)
cv_binom = GLMNet.glmnetcv(X, y, GLMNet.Binomial(), alpha=0.0)

The final line of code produces the following error:

no method glmnet!(Array{Any,1},Array{Float64,2},Array{Float64,1},Binomial)
at In[100]:3
 in glmnet at /home/dhimmels/.julia/v0.2/GLMNet/src/GLMNet.jl:346
 in glmnetcv at /home/dhimmels/.julia/v0.2/GLMNet/src/GLMNet.jl:382

I am new to julia, so perhaps I'm missing something. Thanks.

[simonster]

From the docs:

For logistic models, y is a m x 2 matrix, where the first column is the count of negative responses for each row in X and the second column is the count of positive responses.

but ideally we'd support y given as a vector (see #3)

[dhimmel]

Converting my y vector to the matrix did the trick:

y_mat = hcat(abs(y - 1), y)
cv_binom = GLMNet.glmnetcv(X, y_mat, GLMNet.Binomial(), alpha=0.0)

In the past when using cv.glment in R, I've used the lambda.1se (largest lambda at which the MSE is within one standard error of the minimal MSE) for lambda. Is there any way to easily retrieve this value, see the coefficients for this lambda, and make predictions using that model?

Thanks,
Daniel

[IainNZ]

@simonster I'd be interested about the lambda.1se thing too, if you know.

[simonster]

I would guess lambda.1se is something like:

minloss, index = findmin(cv.meanloss)
max(cv.lambda[cv.meanloss .<= minloss+cv.stdloss[index]])

but I'll double-check with the R interface tomorrow.

[simonster]

This is the R code that computes lambda.1se. What I wrote above seems right. I'll include this when I get around to making a general cross validation framework for my various Lasso packages.

[dhimmel]

Great, thanks for the interim solution to calculate lambda using the 'one-standard-error' rule.