Posterior Optimal E-Values for Logistic Regression

We investigate logistic regression, and ask how to use e-values to distinguish which features are important for a binary classification problem. We adapt an approach based on projecting the posterior distribution on the null. We give a method to compute the projection with the Frank-Wolfe algorithm, by introducing batches of data and giving the resulting convergence rate of the the algorithm in this case. Finally, we give experiments to evaluate our method's performance depending on hyperparameters and compare it to existing method, and we showcase our method a real dataset.