Databases Reference
In-Depth Information
observations are positive and generate revenue. At some point, the proportion of
positive treated observations relative to the total treated begins to fall. The $222
cost of treatment is applied to all, but there are fewer and fewer positives to
bring in additional revenue. Eventually the expected net revenue gain drops
below zero and continues downward as more and more non-buyers are offered
the promotion.
In looking at the chart, the gain using the validation data appears to peak
somewhere around 10% selected. You might ask why it would continue to
increase after passing the 8.2% buyer level in the dataset. The answer is that
when the top 8.2% of the observations are selected, some will be non-buyers
due to misclassifications by the classifier. As the percentage treated is
increased, there will still be some buyers selected. This will continue to
boost the revenue gain until the increased cost of non-buyers wipes out
any gains by the newly selected buyers. This appears to happen at about the
10% level.
Compare the profit analysis based on your ann model with a profit analysis
using the decision tree model. How do they compare?
Classification Ensembles
Given the three classifiers modeling the same dataset, one might wonder if a
combination of the three could collectively generate better predictions than any
single model. In VisMiner, a combination of models is referred to as an
ensemble
.
Right-click on the “Tree Classifier: Target” model.
Select “Create classifier ensemble”.
A new classification model is created with the name “Ensemble: Target”. The
single arrow from the tree classifier model to the ensemble model indicates that
the tree classifier is the only participant in this newly created ensemble.
Drag the “ANN Classifier: Target” model over the ensemble and drop.
With a second model added to the ensemble, VisMiner uses probability
predictions from each model to generate a single combined prediction.
Hover over the “Ensemble: Target” model to see its performance measures.
Hover over the two contributing models to compare their performance
measures with those of the ensemble.
