4.1 Synthetic dataset is shown in Figure 4.1(a) with normal instances
(blue), global anomalies (yellow), and local anomalies (red and
green). Figure 4.1(b) LOF score with k = 10. 4.1(c) IForest score.
The anomalousness are visualized as height bar over all the
instances. For each algorithm output, the anomalousness scores are
normalized in the range of [0, 1] to have an easy comparison. We
can see that both LOF and IForest fail to totally distinguish local
anomalies from normal instances. . . . . . . . . . . . . . . . . . . 65
4.2 Histogram of anomalies (red) and normal instances (blue) on the
first four eigenvectors 1 of ionosphere dataset (a popular benchmark
dataset for anomaly detection [96] [63] [110]). Some anomalies
have overlapped distribution with parts of normal instances and
therefore it is nontrivial to separate them simply by difference between
attribute distributions. . . . . . . . . . . . . . . . . . . . . . 67