General approach for churn prediction with genetic algorithm optimized k-nearest neighbor framework

Abstract

Customer churn has become one of the most significant topics in today‟s business. It has become a major challenge for a business with the evolving market and low barriers to switch between the service providers. It has identified that, retaining the old customers is more profitable for a company than acquiring new customers. That motivates business personnel for churn prediction. Service providers can get necessary measures to retain their customers if they could gain prior knowledge on the probable churns in the customer base. But, churn prediction is considered a difficult task. Various attempts have been made in predicting churn and churn related information. Different data mining techniques had been used in developing churn models. Regression analysis, decision tree based methods and neural network based methods were among the most commonly used techniques. The most successful models suffered from low interpretability which is a main consideration in a churn model while some of the models were domain specific. K nearest neighbor classifier is one of the best algorithms to be used in classifications. But it has been rarely used in churn prediction. Genetic algorithms are considered an optimization technique which could be used in optimizing performance of other algorithms. Genetic Algorithm Optimized K Nearest Neighbor (gaKnn) is a framework that has tested for its high accuracy. Hence, we developed a Tool based on the gaKnn framework which could be used for churn prediction. We also incorporated two voting mechanisms; Bayesian weights and class confidence weights (ccw) to weight the prediction in order to address misclassification issues occur due to class skew.