Java Reference
In-Depth Information
first step. Once we know which customers are high value, we can
take steps to encourage them to remain customers, especially if
we've identified them as likely to leave. Just as important as knowing
which customers to keep is knowing which customers to let go. As
noted above, low-value customers can be costly to maintain. For
example, if a customer frequently uses your call center with ques-
tions or problems yet buys minimal product or services, it may be
more cost effective to allow such customers to leave. Both active and
passive actions can help to reduce low-value customers. Combining
data mining results with rule-based systems can help to automati-
cally recommend actions for certain customers.
For customers who are likely to leave, businesses need to under-
stand if there are any common factors (e.g., a certain age or ethnic
group, geographic region, or type of products sold) that are common
among dissatisfied customers. If such factors are identified, busi-
nesses may be able to take more effective actions to avert a
customer's decision to leave. Data mining can be used to identify the
factors that play most heavily in determining an outcome. The data
mining technique attribute importance can help here (see Section 4.4)
as well as the decision tree algorithm which produces rules that high-
light the specific attribute values that result in a dissatisfied
customer. For example, a decision tree rule may indicate that high-
income, 30-something female customers from the Northeast who
purchased the latest product offering and had an unsuccessful call
center experience cancelled their service.
Knowing in aggregate the number of customers likely to leave,
say in the next three months, can give a manager a reasonable
estimate for resource allocation, either in the number of support staff
needed to try to retain these customers or budget to provide incen-
tives to customers at risk. In this case, classification techniques allow
the building of models that generate a probability of each customer
to leave. Summing these probabilities provides the estimate of total
attrition.
2.1.3
Response Modeling
The essence of response modeling is to determine whether or not
someone will respond positively to a request or offer. That request
may be to purchase a product, complete a survey, donate money, or
participate in a clinical trial. The motivations for response modeling
are simple: reduce costs by soliciting fewer people who have a
greater likelihood of responding, increase return on investment by
Search WWH ::




Custom Search