Solving Problems in Industry - Java Data Mining: Strategy, Standard, and Practice - page 35

on sale that uses a steering wheel input device. If the retailer does not
foresee the relationship, there may be plenty of computer games in
stock (what the retailer planned to promote), but customers may not
be able to buy the complementary steering wheel input device since
the product may have sold out prematurely. As a result, the retailer
loses additional sales from customers who want that input device, or
even loses the sale of the computer game from customers wanting
both items and purchasing none.

2.1.6 New Product Line Development

Answering which bundle of products should be offered to custom-
ers, perhaps with pricing discounts, is a challenging task. Knowing
which products are normally purchased together, possibly out of
thousands, is daunting. Since not all customers behave the same, a
bundling strategy should start with groups of similar customers. The
first step is to identify a set of customer segments or target markets.
Consider a typical customer database, including individuals with
a wide range of ages, incomes, and interests. Using clustering
techniques, we can automatically identify customer segments. These
segments might include such groups as “wealthy seniors who enjoy
the outdoors,” “young families in the suburbs,” “college students
who rent,” and “large families in cities who use public transporta-
tion.” We then partition the market basket data associated with each
customer according to the customer's assigned segment. The next
step is to use an association technique to find those products that are
frequently purchased together within each segment. For example, we
may find that “young families in the suburbs” may purchase several
life insurance policies on family members as well as home insurance,
but not car insurance. These now form the candidate product bundles
that can be targeted for each of the customer groups. In the example,
this could lead to a product bundle for multiple life insurance policies
or a more complete bundle that includes car insurance to get this
customer segment to purchase car insurance from this insurance
company as well. The basic process is depicted in Figure 2-6.
Usually it is not enough to just identify possible product bundles.
It is important to understand how likely a customer group is to pur-
chase the new product bundle, and how much profit we can expect
to make from such customers. To determine the success of a product
bundle, a test marketing campaign can be performed, much like the
response modeling problem of Section 2.1.3. From this we can deter-
mine not only if a significant percentage of customers will purchase