1
$\begingroup$

I have trained and XGBoost by enforcing no-feaure interaction and calculated Global Shap values:

enter image description here

It looks like only 6 features have some SHAP values, whilst the remaining ones have a SHAP value of 0.

Question. If a feature has a SHAP value of 0 across all records in the sample, does it mean that that feature has not been included in the model?

$\endgroup$
1
  • $\begingroup$ you can check the importance gain or number of splits to see if a feature is actually used in the model. $\endgroup$ Apr 26, 2023 at 13:30

1 Answer 1

1
$\begingroup$

I am in the same situation, the analyzes carried out by me so far are:

  1. Missing data can influence: Studying my data I did not find a direct relationship, since I found the same compartment, for example: with 6 nans of features of a total of 19 v/s 0 nans of features of a total of 19 and both give me all its features except for the clear bias 0.
  2. Training my model: in general the training performs well around 70% in most of the recall metrics, f1_score, etc.
  3. Different ways to get the shap values: currently I use the same model to get it model.get_booster().predict(DM_data, pred_contribs=True) or else model.get_booster().predict(DM_data, pred_interactios=True) and I get the same result in many cases.
  4. Change the min_child_weight value: I read somewhere that the weight of the leaf in the tree can influence, but there was no difference either. Grades: My problem is unbalanced and from what I saw my bias around -1.7 indicates that it is biased to one class and it makes sense. I don't know if it has a direct relationship, but I'll tell you anyway.

Conclusion: I still can't find and I'm investigating the reason for the problem, my hasty conclusion would be that the model is not capable of achieving an interpretation with the given data and is only predicting thanks to the bias of the model.

I hope we can find the truth. Greetings

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.