|Your Navigation Path||Case Base
Call center chain
A telecommunications company introduces a new service and prepares a marketing campaign for it. To make the marketing campaign more efficient, the company relies on its available data about its customers: a subset of customers that are most likely to subscribe to the new service offer is identified using data mining. Customer's profiles are built on the basis of billing data (from existing subscriptions), and data from a call center from where previous marketing campaigns were led. A classifier (here, a decision tree) is trained to distinguish between customers that have previously been positive about new products and technical developments, and other customers. The trained classifier is then applied to a new group of clients with unknown profile. This allows to direct the marketing campaign to such clients as are more likely to respond positively.
The telecommunications company has used three sources of data for this Case: Call Detail Records (Concept CallDetails: length, number called etc. for each telephone call a client makes); the list of current customers (Concept ActualClients); and results of previous call center examinations of some customers (Concept CCExamination). While the MiningMart Case model was developed using artificially generated data for confidentiality reasons, the exact format of the original data was used. After preprocessing, the Concept DataMiningInputConcept is used for mining.
The main overall goal was to lower marketing costs by leading a well-targeted marketing campaign, directed at the group of clients which are most likely to give a positive answer to an offer. To plan this, a deeper knowledge of customers was needed, and was a goal in itself. The more direct aim, then, was to increase the percentage of positive responses by targeted customers, with respect to the percentage resulting from untargeted campaigns. Customer profiles built from the data proved to be useful to increase this percentage, and can be re-used for future applications.
The preprocessing of the data included the computation of various statistical features for each client that reflect this client's history of telecommunication behaviour. This information was then joined with the data from previous call center interviews with the clients. The result is a single attribute that distinguishes customers into four classes: not interested, interested in offer, already subscriber, unknown. A decision tree is then learned that can seperate further customers according to these classes.
Janusz Granat, National Institute of Telecommunications (NIT), Poland; Telephone: +48-22 512 8303; Email: email@example.com; Website: http://www.itl.waw.pl