Divergence from GLMnet when using a matrix with many variables about lasso.jl HOT 2 OPEN

The packages may be generating a different path of regularization lambdas.
You can get them from the lasso path as l.λ, and from GLMNet as g.lambda.

Also, in your example it looks like you are picking the last coefs of the regularization path, but those are not necessarily the most interesting ones. Take a look at the docs here.

from lasso.jl.

Hi, thanks for chipping in.

I don't think the regularization lambdas is where the differences are coming from. In the maxnet algorithm the lambdas are generated inside the algorithm and not by the packages. Even if I force the lambdas to be identical, I see the same kind of behaviour.

The same goes for when I look at some other part of the regularization path (maxnet always takes the last one, but I see your point).

E.g. in this example I take coefficients halfway in the path and force the lambdas to be identical, and lasso_glmnet_dif(1000, 1000, 5) is still around 0.02.

That seems like a big difference for it to come from floating point errors, which leads me to think the algorithms are somehow different?

function lasso_glmnet_dif(nrow, ncol, n_col_contributing)
    data = rand(nrow, ncol)
    outcome = mean(data[:, 1:n_col_contributing], dims = 1)[:,1] .> rand(nrow)
    presence_matrix = [1 .- outcome outcome]

    l = Lasso.fit(LassoPath, data, outcome, Binomial())
    g = GLMNet.glmnet(data, presence_matrix, Binomial(); lambda = l.λ)

    lcoefs = Vector(l.coefs[:,floor(Int,end/2)])
    gcoefs = g.betas[:, floor(Int,end/2)]

    mean(abs, lcoefs .- gcoefs)
end

from lasso.jl.