SUPORT VECTOR MACHINES FOR CHURN PREDICTION IN THE MOBILE TELECOMMUNICATIONS INDUSTRY

This book presents a CRISP-DM data mining project for implementing a classification model that achieves a predictive performance very close to the ideal model, namely of 99.70%.


This model yields such a high accuracy, mainly, due to the proprietary architecture of the machine learning algorithm used. We implement a support vector machine which is improved using multiple techniques existent in the literature. A detailed theoretical explanation is offered regarding support vector machines, learning algorithms and several optimization algorithms, and each decision taken in building the final architecture is motivated.


To demonstrate the predictive performance of our classification model, we use a telecommunications synthetic dataset that contains call details records (CDR) for 3,333 customers, with 21 independent variables and one dependent variable which indicates the past behavior of these customers with respect to churn. This is a generic dataset frequently used in research as a benchmark for testing different architectures of machine learning algorithms proposed for classification.


The methodology presented in this book is scalable to datasets that have hundreds of thousands of instances and hundreds or thousands of variables coming from various industries such as telecommunications, finance, astronomy, biotech, marketing, healthcare, and many others, and can be applied to any real world classification problem.