|Your Navigation Path||Case Base
Part 1 - Treat missing values in CDR
Part 2 - Transpose CDR from transactional to relational form
Part 3 - Transpose REVENUES from transactional to relational form
Part 4 - Create derived attributes and customer profile
Part 5 - Churn Modeling
This Case was developed in the telecommunications industry. The aim was to understand when and why a company's customers are likely to leave the customer base, so that remedial action where appropriate can be planned and taken. Customers become ''churners'' when they discontinue their subscription and move their business to a competitor. One solution to combating churn in telecommunications industries is to use data mining techniques. Data mining may be used in churn analysis to perform two key tasks: a) predict whether a particular customer will churn and when it will happen; and b) understand why particular customers churn. By predicting which customers are likely to churn, the telecommunications company can reduce the rate of churn by offering the customer new incentives to stay. By understanding why customers churn, the company can also work on changing their service so as to satisfy these customers ahead of time. In addition, the chance of the customer churning after action is taken can be assessed by the data mining tool so as to choose the best strategy in terms of cost and effort. Churn management has emerged as a crucial competitive weapon, and a foundation for an entire range of customer-focused marketing efforts. Part of the process is determining customer value, as the company may prefer to let unprofitable customers go. When this kind of customer profile is available to a company, marketing managers may take informed and strategic action to minimize defections, win back valued defectors, and attract more cost-effectively the right kind of customers in the future, including those that are least likely to churn.
The telecommunications company that developed this Case used four different sources of data as follows. Call Detail Records: For every telephone call a client makes, all information such as length, number called etc. is stored in a Call Detail Records table. Customer demographics: Basic customer information like age and gender. Revenues: Basic billing data in transactional form. Each record provides the revenue (difference between income and costs) related to each user aggregated on a monthly basis over five months. Services: Basic information about the type of service a customer has subscribed to, such as: handset (cell-phone model/class), length of service (contract duration in months), number of dropped calls, or tariff plan and tariff type used for billing. This last table also includes the target attribute, that is, whether or not the customer has churned in the sixth month.
The main goal in this Case was to understand why particular customers churn in order to find appropriate strategies of customer retention. More technically, the aim was to minimize the prediction error of the learned classifier that separates customers according to whether they have churned in the sixth month or not. A classifier with a small prediction error can then be applied to further customers; those who are classified as likely to churn can subsequently become the target of specific marketing actions. See also Overview.
As the classifier, a decision tree was learned using the C4.5 algorithm. The customers' telecommunication behaviour information from five months was used together with general customer information to predict the churn state in the sixth month. The preprocessing of data follows the idea of separating customers according to their value to the company, as measured by the revenue (difference between income and costs). Preprocessing has been divided into five tasks as follows: a) handling missing values in CDR (call detail records); b) transpose CDR from transactional form to relational form; c) transpose REVENUES from transactional to relational form; d) create derived attributes and customer profile; and finally e) the data mining step, that is, churn prediction modelling. For each of these tasks a Chain containing several Steps has been developed.
Marco Richeldi, Department of Economics and Regulatory Affairs, Telecom Italia Lab S.p.a; Telephone: +39-011-2288073; Email: firstname.lastname@example.org; Website: www.tilab.com