We should return the average of magnitudes of shap values per feature over the complete training dataset as a new result type. This will allow a global ranking features by their importance for the complete dataset.