Shap: get SHAP value for each feature

Created on 24 Dec 2019 · 2Comments · Source: slundberg/shap

Is it possible to the SHAP value for each feature?

Source

alexlaurence

Most helpful comment

From the docstrings of the shap_values method, we can see that:

For models with a single output this returns a matrix of SHAP values 
(# samples x # features)

So unless you have a different situation/case, you should be getting # samples arrays with SHAP values for each feature when calling this method.
Do let me know if this does not work for you.

ibuda on 27 Dec 2019

👍2

All 2 comments

From the docstrings of the shap_values method, we can see that:

For models with a single output this returns a matrix of SHAP values 
(# samples x # features)

So unless you have a different situation/case, you should be getting # samples arrays with SHAP values for each feature when calling this method.
Do let me know if this does not work for you.

ibuda on 27 Dec 2019

👍2

Unless I'm mistaken, you can simply average over the SHAP values for each sample for each feature. e.g. here's a snippet:
mean_shap_feature_values = pd.DataFrame(shap_values, columns=your_xs_dataframe.columns).abs().mean(axis=0).sort_values(ascending=False)
These should align with your simple SHAP importance plot (summary_plot with plot_type='bar').

It would be a nice convenience function for SHAP to expose these "importances" with an API like SKL's rf.feature_importances_ or similar.