Databases Reference
In-Depth Information
class A
class B
Y
Slack
X
Figure 4.12
Slack Requirement
During the search, to balance the contribution of the two objectives, the slack
total is adjusted by a cost factor (
C
). Such that:
m
i¼
1
z
i
slack contribution ¼ C
where
z
i
is the slack requirement of the
i
th
observation, recognizing that most will be
zero
m
is the number of observations.
The value of
C
is set by the analyst prior to algorithm execution.
Data transformations
Look at Figure 4.13a. Attempting to find a dividing line for this dataset would be
futile. However, suppose that instead of submitting (X, Y) observation pairs to
the algorithm, we submit (X
2
,Y
2
), as plotted in Figure 4.13b. The SVM will
have no trouble finding a dividing line for this set. The point here is that a
transformation of the data prior to processing by the algorithm may generate
classification results not achievable by the original data.
Most implementations of SVM algorithms include a set of potential
transformations from which the analyst may choose. These are known as