Claim 3.4 fitted function is closer to being additive

[do8572]

Fig 7 investigates the SHAP values for an RF trained on the breast-cancer dataset. After applying HS, the SHAP values for each feature have tighter clusters. Each cluster comprises a group of samples for which the feature contributes a similar amount to the predicted response, and hence can be interpreted without taking into account feature interactions, thereby reducing cognitive burden. More globally, the clustering effect suggests that HS improves prediction performance by regularizing some unnecessary interactions in the model, making the fitted function closer to being additive. Appendix S5.2 shows the SHAP plots for all 8 classification datasets, showing that HS consistently leads to more clustered SHAP values.

[dmohorcic]

This is part of Claim 3.3 #13, and I have no idea how to approach this. @do8572 any thoughts?