Interpret's LIME is not the same as original LIME

[Dola47]

Dear All,

First Question:

I use the original LIME to explain some instances of my dataset. I decided to use Interpret's LIME to explain the same instances due to the better user interface we have from Interpret. However, I am explaining the same instance, and using the same conditions the results are different. Any ideas about why the results are not the same?

Here are the results from LIME:

Here are the results from Interpret LIME:

Some thoughts from my side:
1- Is Interpret's LIME using random seed as in LIME or not, if you are not using it then the results will be random from call to call?
2- Is Interpret's LIME by default is doing classification or not?
3- Is Interpret's LIME importing by default the default values set to the different parameters in LimeTabularExplainer and LIME's explain instance or I have to specify them?

This is how I called Interpret's LIME to be exactly under the same conditions as the LIME explainer I am using:

lime_trial = LimeTabular(predict_fn=model.explainer._build_predict_fn(), data=label_encoded_data_train_x, training_labels=label_encoded_data_train_y, categorical_features=model.explainer.categorical_indices, categorical_names=model.explainer.categorical_names, feature_selection='none', feature_names=model.explainer.features_for_lime, discretize_continuous=True, discretizer='quartile', class_names = ['NOK', 'OK'], kernel_width = None, verbose=False, explain_kwargs={'num_features': 10, 'num_samples': 5000, 'top_labels': None, 'distance_metric': 'euclidean', 'model_regressor': None})

lime_trial_xgb = lime_trial.explain_local(label_encoded_interesting_parts_x, label_encoded_interesting_parts_y, name='LIME')

Second Question:

For SHAP kernel explainer, I am not sure if I passed the nsamples parameter used in SHAP during calling the shape values, to the ShapKernel in Interpret's it will work probably or not. I see that It does not make any difference when I pass it or not, so it appears that the class is not importing the parameter correctly.

Third Question:

I know that the ebm model deals with the categorical features automatically is there a way to stop that? I prefer to process my features on my own then fit the model to the processed features directly.

Many thanks.