4.4 70 nearest neighbors (in green) of red instance with GAU (Figure
4.4(a)) and AGK (Figure 4.4(b)), which shows that AGK has better
manifold-aware property than GAU. . . . . . . . . . . . . . . . . . 73
4.5 Illustration of LAD (Local Anomaly Descriptor, Equation 4.10)
which calculates weighted average of neighbor differences. It is
one of the ways to take the neighborhood distribution into consideration
[135]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.6 LAD with large t fails to reveal the local anomalousness 2
(Figure
4.6(a)) due to the over-diffusion. Comparably, FDD acts robustly
in measuring anomalousness regardless of small or large scaling
parameter (Figure 4.6(b) and 4.6(c)). . . . . . . . . . . . . . . . . . 79