Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Numerical stability warning #522

Open
olivierlabayle opened this issue Mar 27, 2023 · 2 comments
Open

Numerical stability warning #522

olivierlabayle opened this issue Mar 27, 2023 · 2 comments

Comments

@olivierlabayle
Copy link

olivierlabayle commented Mar 27, 2023

log_loss_pb_dataset.csv
Hi,

This may be a bit of a long shot. I've noticed that on a specific dataset (attached) if I run a GLM with both GLM.jl and the R glm library, R will complain about: "glm.fit: fitted probabilities numerically 0 or 1 occurred", but GLM.jl will not. Is there anything particular which is done here to solve that issue, or is it just ignored?

Julia code:

using CSV, GLM, DataFrames

data = CSV.read("log_loss_pb_dataset.csv", DataFrame)
X = reshape(data[!, :covariate], size(data,1), 1)
f = @formula(y ~ 0 + covariate)
glm(f, data, Bernoulli(), offset=data.offset)

R code:

data <- read.csv(
    file = "/Users/olivierlabayle/Dev/TMLE.jl/test/log_loss_pb_dataset.csv",
    stringsAsFactors = TRUE
    )
data$y <- as.numeric(data$y) - 1
res <- glm(y ~ 0 + covariate, family = binomial(), data, offset=data$offset)
@nalimilan
Copy link
Member

Some predicted values are indeed equal to 1 and some very close to zero. But that doesn't always indicate a problem. R prints a warning even though the model can sometimes be interpreted just fine, while in GLM.jl we tend to avoid printing warnings. Instead, if there's a problem you should be able to spot them due to very large standard errors on problematic variables I think. You can find more details by searching for the R error message in a search engine.

@olivierlabayle
Copy link
Author

Thank you for your reply, I think I've made the situation a bit more clear now. Unless I'm missing something, I think the following code illustrates how the logloss reached by a GLM (in this specific example) is way worse than that of a constant classifier.

This is using MLJ for convenience as I wasn't sure how to extract the pdf from a glm model.

using CSV
using DataFrames
using MLJBase
using MLJModels
using MLJGLMInterface

data = CSV.read("glm_pb_dataset.csv", DataFrame)
X = data[!, [:covariate, :offset]]
y = categorical(data.y)

# Dummy mean classifier logloss
mach_const = machine(
    ConstantClassifier(),
    X,
    y
)
fit!(mach_const)
ll_const = mean(log_loss(MLJBase.predict(mach_const), y))

# Fit using MLJGLMInterface
mach_glm = machine(
    LinearBinaryClassifier(fit_intercept=false, offsetcol=:offset),
    X,
    y
)
fit!(mach_glm)
ll_glm = mean(log_loss(MLJBase.predict(mach_glm), y))

# What would be the logloss if the coefficient is set to 0?
mach_glm.fitresult[1].coefs[1] = 0.0
ll_glm_0 = mean(log_loss(MLJBase.predict(mach_glm), y))

println(ll_const)
println(ll_glm)
println(ll_glm_0)

Output:

0.6904066761889491
17.256097466399517
0.22942599075747547

I also join the dataset:

glm_pb_dataset.csv

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants