Claim 3.1 removes sampling artifacts & simplifies boundaries (Fig. 5)

[do8572]

Fig 5 shows an example of simplified decision boundaries. On the diabetes dataset (Smith et al., 1988), RF can achieve strong performance (AUC 0.733) even when fitted to only two features. When HS is applied to this RF, the performance increases (to an AUC of 0.787), but the decision boundary also becomes considerably smoother and less fragmented. Since these two features are the only inputs to the model, these smooth boundaries enable a user to identify much clearer regions for high-risk predictions. Appendix S5.1 plots decision boundaries for all 8 classification datasets, showing that HS consistently makes boundaries smoother.9