Page 50 - Machine Learning for Subsurface Characterization

FIG. 1.D1 ROC curves for the isolation forest applied on the subset (A) FS2 and (C) FS2* of
Dataset #1. PR curves for the isolation forest applied on the subset (B) FS2 and (D) FS2* of Dataset
#1. ROC curve indicates a great performance of the isolation forest model, but PR curve indicates
there is room for improvement especially in FS2 dataset. PR curve indicates that the performance on
FS2** is better than FS2, which is aligned with other evaluation metric. Compared with ROC curve,
PR curve is especially suitable when there is outlier-inlier imbalance in the dataset.

FIG. 1.D2 ROC curves for the one-class SVM applied on the subset (A) FS2 and (C) FS2* of
Dataset #1. PR curves for the one-class SVM applied on the subset (B) FS2 and (D) FS2* of Dataset
#1. ROC curve indicates similar performances of the isolation forest and one-class SVM, but PR
curve indicates the performance of one-class SVM is much better than isolation forest on Dataset
#1. Compared with ROC curve, PR curve is especially suitable when there is outlier-inlier imbalance
in the dataset.