Getting Started - Java Data Mining: Strategy, Standard, and Practice - page 124

of the likely responders from only 40 percent of the data. The
DMWhizz data miner may decide to specify other algorithms besides
the decision tree that the DME selected, change some of the decision
tree algorithm settings, or prepare the data differently to see if a better
lift can be achieved. If another model produced a higher lift for 40 per-
cent of the customers, perhaps 0.7, the data miner would likely choose
that model. If another model produced the lift of 0.6 at 35 percent of
the customers, DMWhizz may choose to send the campaign to fewer
customers while maintaining the same number of likely responses.
We now move on to the evaluation phase of the process.

6.5 Evaluation

In the evaluation phase, we are interested in understanding how well
the model meets the business objectives. We see from the test results of
the model produced above that the lift is .6. How does this meet our
business objectives? Out of 1 million customers, historically we know
that 3 percent, or 30,000, should respond. The lift results tells us that by
contacting the right 400,000 customers (or 40 percent of the 1 million
customer base), DMWhizz can get 60 percent of the likely responders,
or 18,000 (60 percent of 30,000). This is a response rate of 4.5 percent.
Given that DMWhizz's original requirement was to increase the
response rate to 4 percent, the expected 4.5 percent provided by data
mining yields a comfortable margin. As a result, DMWhizz decides
to use this model to score the remaining 980,000 customers and pro-
ceed with the campaign.

6.6 Deployment

In the evaluation phase, DMWhizz found that the model meets the
objective and can be used to complete the campaign. Since we have
already contacted 20,000 of the 1 million customers for our sample
campaign, we apply the model to the remaining 980,000 customer
cases and send a mailing to the top 40 percent of customers predicted
to respond to the campaign. We use the code below to score these
customers in batch, that is, all at once, and produce a separate table
that includes the customer identifier and the probability that the
customer will respond. All the cases are ordered by their probability
in descending order.
To apply the model, we first create a physical dataset object from
the CUSTOMER_APPLY table. The CUSTOMER_APPLY table contains