Solving Problems in Industry - Java Data Mining: Strategy, Standard, and Practice - page 38

is computed manually. This score is combined with customer
demographic and other data as noted above. If we were interested in
classifying customers as a high, medium, or low credit risk we could
use a classification technique. If we wanted to predict a numerical
score, we could use a regression technique that predicts a continuous
numerical value.
Another approach uses historical data on customers who either
failed or succeeded in fulfilling their loan obligations in the past.
Instead of using manually computed “scores,” a classification
algorithm learns to predict the probability of default on a loan. After
these probabilities have been established, a scale can be introduced
for ranking customers (assigning a credit score), usually based on a
desired distribution of scores (e.g., 5% must be in the top range—
“AAA” rating according to some classifications; 25% in the next
range, and so on).

2.1.9 Warranty Analysis

Anyone who buys products has likely had some of those products
break. Some of those products will be under warranty, which
means that someone—the retailer, manufacturer, or independent
warranter—will make certain repairs free of charge. Many products
come with warranties automatically, for example, you will see “the
manufacturer warranties this product to be free from defects for a
period of 90 days from purchase.” More common today are cus-
tomer-purchased, or extended, warranties. Although extended
warranties are often regarded as a “cash cow” for retailers, depend-
ing on the industry, the costs associated with servicing products
are, to some extent, a gamble. The warranter expects a certain per-
centage of products to fail within the warranty period, and builds
the cost for that into either the product cost or the cost of the
extended warranty. By using advanced analytic techniques, war-
ranters can better manage the seemingly unpredictable and uncon-
trollable expenses associated with warranties. Manufacturers and
retailers need good algorithms for predicting future claims, very
reliable products, or dramatically overpriced warranty offerings.
To reduce internal warranty costs where multiple suppliers,
assemblers, dealers, and repair centers are involved, it is important
to understand where product failures originate to improve cost
recovery and processes. Because data mining can attribute problems
to the various parties in the product path, data mining can be used to
identify the root causes behind warranty repairs, for example, a