Page 36 - Machine Learning for Subsurface Characterization

Unsupervised outlier detection techniques Chapter 1 21

FIG. 1.5 Three of the seven 3D scatterplots containing various combinations of three (A–C) out of
the seven available logs in the offshore dataset. DBSCAN clustering was optimized and sequentially
7
applied on each of the combination of 3 logs out of the total 35 possible combinations ( C 3 ) to iden-
tify outliers in the offshore dataset.

points (dark gray in the print version) show the known inliers, and the red points
(light gray in the print version) show the known outliers, as detected by the
DBSCAN model. The model has been tuned to where the main body of normal
data is all blue points (dark gray in the print version) while all the points that
deviate from this main body are red (light gray in the print version).

4.4 Metrics/scores for the assessment of the performances of
unsupervised ODTs on the conventional logs
In real-world implementations, there is no means and metrics to evaluate the
performance of the unsupervised ODTs. However, for purposes of comparative
study of the four unsupervised ODTs as planned in this chapter, unsupervised
ODTs process the four previously mentioned validation datasets, namely
Datasets #1 to #4, to assign a label (either outlier or inlier) to each depth (sam-
ple) in the dataset. A label is assigned to each depth based on the log responses
(feature vector) at that depth. In real-world applications of unsupervised outlier
detection, there is no prior information of outliers, and no such inlier vs. outlier
labels are present. For purposes of comparative study of the performances of the
unsupervised ODTs, we created the four datasets, namely, Datasets #1, #2, #3,
and #4 containing well-defined, manually verified outlier vs. inlier labels.
Therefore, for evaluating the unsupervised ODTs, we compare the labels known
a priori against the new labels assigned by the unsupervised ODT using metrics/
scores popularly used to evaluate the classification methods. In a real-world