Databases Reference
In-Depth Information
class A
class B
Y
X
Figure 4.3
Classification Plot
A final advantage is that the danger of overfit by the algorithm is less than for
other classification algorithms. This is especially true for training implementa-
tions that have complete split termination rules and use a validation dataset to
guide check for overfitting.
Limitations
Because decision tree algorithms consider only one input attribute at a time,
they quite frequently perform worse than other classification algorithms. For
example, consider the scatter plot of input attributes X and Y in Figure 4.3. A
decision tree algorithmwould try to split the observations based on either X or Y
alone. The algorithm would not be able to locate the dividing line so visibly
obvious. Other algorithms such as artificial neural networks and support vector
machines would have no trouble finding and implementing the split.
Another issue is with respect to the robustness of the modeler. For example,
if two input attributes are nearly equal in their discriminatory ability at the first
level split, splitting with the slightly better attribute may generate a totally
different tree than if the other were chosen. In fact, the overall performance of
the resulting tree may be better if the second best attribute is chosen for the first
split rather than the best.
Artificial Neural Networks
Artificial neural networks
(ANNs) are computing methodologies designed to
mimic functions within the brain, giving computers the ability to learn from
