A Novel Genetic Algorithm Based Method for Building Accurate and Comprehensible Churn Prediction Models

Customer churn has become a critical problem for all companies inparticular for those that are operating in service-based industries suchas telecommunication industry. data mining techniques have beenused for constructing churn prediction models. past research in churnprediction context have mainly focused on the accuracy aspect of theconstructed churn models. however, in addition to the accuracy,comprehensibility aspect should be considered in evaluating a churnprediction model. being comprehensible, a model can reveal the mainreasons for customer churn; thereby mangers can use such informationfor effective decisions making about marketing actions. in this paper,we demonstrate the application of a genetic-algorithm (ga) methodfor building accurate and comprehensible churn prediction model. theproposed method, ga-based method uses a wrapper based featureselection approach for choosing the best feature subset. the keyadvantage of this method, is taking into account the comprehensibilitymeasure (measured as the number of rules extracted from c4.5decision tree) in evaluating the performance of a candidate model. thega-based method is compared to the two filter feature selectionmethods including chi-squared based and correlation based featureselection using two telecommunication churn datasets. the results ofexperiments indicated that the ga-based method performs better thanthe two filter methods in terms of both accuracy and comprehensibility