Databases Reference
In-Depth Information
Figure 5.14
Ensemble Weights
Although it won't always be the case, ensembles can produce results that
outperform any of the contributing models.
By default, each model contributes equally to the class prediction. This can be
overridden by the user.
Right-click on the “Ensemble: Target” model.
Select “Edit ensemble weights”.
Drag the dividing line between the red (tree classifier) and blue (ANN
classifier) to the left until the tree classifier reaches the 25% weight level
(Figure 5.14).
Click “Save”.
Again hover over the “Ensemble: Target” model to review the performance
measures.
Each time the weights of an ensemble are adjusted, VisMiner recomputes the
performance measures. The weight distribution that maximizes a given per-
formance measure is never a given. It will vary depending on the participating
models and dataset. Obviously, the user can manually search for an optimal
weight setting through a trial and error process. However, VisMiner does
provide an automated search mechanism when there are exactly two contri-
buting models to the ensemble. Access this feature by right-clicking on the
ensemble, then selecting either “Find best (auc) weights” or “Find best (error)
weights” to search the weight distribution that maximizes either the AUC or the
error performance measure respectively.
Model Application
If the only objective in performing data mining is to locate and understand the
relationships between attributes, you will have completed that task using
