I am running a LGBM model for a regression problem and then using 'Shap' package to identify the contribution of the different features.
My problem is that 'Shap' is giving me 'negative' contribution of certain features, which in real-life can NOT be negative. (i.e. The Feature can only have a positive or zero contribution in explaining the target variable).
I wanted to request if anybody has experience in constraining the LGBM model in a way that 'feature contributoins' are either positive or zero ?
Thanks so much for your advise
i have tried running multiple LGBM models with different hyperparameters (max depth, num_leaves, reg_lambda, random_state) to check what combinations leads to non-negative Shap values, but all in vain.
I have also tried multiple different features as input to LGBM models, but all in vain as well.