|
Abstract
|
Customer churn has become a significant problem and is one of the prime challenges that many in the services
industry are facing. While all kinds of churn lead to incur loss, the loss of low-value customers will be
naturally less damaging than the loss of loyal and high-value ones. So companies need to build a churn
prediction model for their high-value customers. In this paper, a two-phase framework for prediction of highvalue customer churn has been proposed. Phase 1 is the identification phase which takes into account socialnetwork based variables of customers in identifying the high-value ones. The data of an identified high-value
customer is used as the input for Phase 2 to prepare the churn prediction model. Data of a major
telecommunication company has been used to implement the framework. The customers were clustered by
using K-means algorithm. After ranking clusters, the top-cluster was selected according to clusters ratings.
The data belonging to the top cluster is used in churn prediction model building phase. In this phase, two
neuro-fuzzy techniques, namely the adaptive neuro-fuzzy inference system (ANFIS) and the locally linear
neuro-fuzzy (LLNF) have been applied together with locally linear model tree (LoLiMoT) learning algorithm
on churn data. A new algorithm has been devised for comparing these methods with the most widely used
neural networks such as multi layer perceptron (MLP) and radial basis function (RBF) networks. Results of
comparison indicate that the neuro-fuzzy techniques perform better than neural network models and they are
a good candidate for churn prediction purposes
|