Segmentation Applications in Banking - Data Mining Techniques in CRM: Inside Customer Segmentation - page 227

• The level of merchant grouping that should be used in the analysis.
• The segmentation population and whether it should cover all customers and
credit cards or only a subset of particular interest.
The answers to these issues and the reasoning behind the selected approach
are presented next.
BUILDING THE MINING DATASET
The required usage data were available in the organization's mining data mart
and MCIF, which constituted the main information sources for the needs of the
project. The recorded data covered all aspects of credit card use including monthly
balances, number and amount of purchases, cash advance transactions, payments
as well as customer demographics, and account details. Purchase transactions
in particular were recorded per merchant code, tracking the purchasing habits
of each customer. In addition, a multilevel grouping hierarchy of merchant
codes was available in a relevant lookup table, enabling purchases per merchant
category/subcategory to be summarized.
One of the things that had to be decided before building the mining dataset
was the categorization level of the merchants. Since the bank's primary objective
was to differentiate customers according to their general purchasing behaviors,
a high-level categorization into 14 merchant categories (to be presented below)
was selected. The use of a more detailed grouping was beyond the scope for the
particular project and was left to a micro-clustering approach in a later phase.
The preparation of the modeling dataset included the summarization of purchase
transactions per merchant group.
Segmentation level was another issue that had to be determined. The selected
data model and the record level of the final input table should be the same as the
selected segmentation level. So, the marketers had to decide whether they would
group together either customers or credit cards.
In general, each credit card customer can have more than one credit card
account and each credit card account can include one primary and a number of
potential add-on cards. In order to get a unified view of each customer, which
would then facilitate their subsequent marketing strategies, the marketers decided
to proceed with a customer level segmentation. Therefore they had to build the
mining dataset at the customer level.
It had also been decided to use six months of data for the segmentation.
A snapshot of the transactional data based on only a few days or even weeks
might just capture a random twist in customer behavior. Since the objective was to
develop a segmentation scheme that would reflect typical behaviors, the approach