ccaribe9 / shapeffects Goto Github PK

Hi,

First of all thank you for publishing the paper and for providing the code with it for reproducibility. While going through the code, I noticed a slight discrepancy (maybe intended) between the description of the algorithm and the actual routine.

In page 12, it says that :

The algorithm is divided into two phases, a preliminary and optional
phase and the main component. In the first preprocessing phase, a random
variable is introduced into the dataset, specifically, a permutation of the most
influential variable according to the mean of the absolute values of the Shapley
values between the original variables.

However, in the function introduce_random_variable, a uniform distribution is sampled from the most influential variable:

sample_train = np.random.uniform(low=random_variable_train.min(), high=random_variable_train.max(), size=(len(X_train),))
sample_val = np.random.uniform(low=random_variable_val.min(), high=random_variable_val.max(), size=(len(X_val),))

Following the paper definition, it should be:

sample_train = np.random.permutation(random_variable_train)
sample_val = np.random.uniform(random_variable_val)

I doubt that it has a big impact on results but I wanted to check with you nonethless whether it is intended or not. Happy to hear your thoughts.

Thanks,
Thomas