Fig. 3

Graphical illustration of how with certain high-dimensional manifolds setting k too high can cause actual outliers to be classified as normal (false negative) and can cause normal points to be classified as outliers (false positive). Using large k approaching n defeats the purpose of the local distance measure, which is to allow the test to use only local regions in ecological distance space and can cause normal reference samples at the extremes of the distributions to be classified as outliers. On the other hand, if k is too small, then it is not robust to subtle variations in the reference group. By default, the CLOUD test sets the neighborhood size to 5% of the size of the reference set