it is conceptually mistake to use log loss for performance comparison

[Sandy4321]

Describe the bug
it is conceptually mistake to use log loss for performance comparison
since
1
lightgbm as all gmb models
( maybe except https://stanfordmlgroup.github.io/projects/ngboost/ NGBoost: Natural Gradient Boosting for Probabilistic Prediction
Tony Duan*, Anand Avati*, Daisy Yi Ding, Sanjay Basu, Andrew Ng, Alejandro Schuler )
are not good in prediction probabilities
2
then same should be carefully tested for deep NN

3
may you change provement that your code is performing , for example
https://github.com/manujosephv/pytorch_tabular/blob/main/examples/PyTorch%20Tabular%20with%20Bank%20Marketing%20Dataset.ipynb
to compare actual confusion matrices pls
I did some attempt
preds_GATE_Full = tabular_model.predict(test)
but what is the way to calculate ground truth predictions from your models

4
where is description of used algorithms , especially for categorical data values preprocessing ?

[Sandy4321]

ok , i found how to calculate performance

[Sandy4321]

so
confusion_matrix for GATE_Full
array([[7765, 220],
[ 653, 405]], dtype=int64)
confusion_matrix for lightgbm
array([[7684, 301],
[ 528, 530]], dtype=int64)

when modle training time much less for lightgbm ( I used cpu runs)
what is about inference time?
what is about ram and cpu/gpu resources use?