Databases Reference
In-Depth Information
Tests for model significance - at this time, there is no published test of
model significance. By significance, we mean: “Can we measure the
probability that the model predictions are anything more than chance
occurrences?” Unlike the well-studied linear regression models, there is
no “null hypothesis” type test. When using ANNs, the only work-around is
to build the model with enough observations that we feel comfortable that
the model actually will predict accurately when applied to new data. The use
of validation data, discussed in Chapter 5, can give us that assurance.
Support Vector Machines
The algorithm for
support vector machines
(SVM) was developed to improve
classifier performance. Look back at the scatter plot of Figure 4.3. If a decision
tree algorithm is applied to this dataset, it will perform poorly. This is because
the algorithm considers input attributes in isolation when searching for the best
split attribute. As a result, the splits are always parallel to one of the attribute
axes. On the other hand, in its simplest form, the SVM computationally locates
that
line that will best split
the observations according to classification
attribute values.
In the search for that dividing line, there are a number of issues that the
algorithm needs to deal with. Consider Figure 4.10 - a slightly modified version
of Figure 4.3. Three candidate dividing lines are drawn separating the red dots
(class A) from the blue dots (class B). Each would work equally well as a
class A
class B
Y
X
Figure 4.10
Potential Dividing Lines
