A New hybrid classification algorithm for predicting customer churn

Abstract

Decision trees, support vector machine and gradient boosting are very popular algorithms for predicting the customer churn with good comprehensibility and strong predictive performance. In spite ofall strengths, the decision trees be likely have some problems forholding linear-relations amongthe variables, support vector machine performs marginally better than logistic regression, and gradient boosting givesgreater results when compared with logistic regression, with less development effort. Hencenew hybrid-algorithm, aboosting leaf model (BLM), was proposed forclassifying the data in better way. The basic idea behind this BLM is diverse models was constructed among the segments of data instead of entire dataset thusleads to improved predictive performances how ever observance comprehensibility among those models which constructed on leaves. ThisBLM resides two stages they are one is segmentation and the other one is prediction stages. Inthe first stageby using decision tree segments of customers are identified and second stagemodel wasappliedon each leaf of the tree. This new hybrid-approach was bench-marked compared with decision trees, support leaf model, andlogit leaf model (LLM)regards predictive performance and comprehensibility. The top decile lift (TDL), area under Receiver Operating Characteristics curve (AUC) which used to measure theirpredictive performancesof which BLM marksknowinglyimprovedtheirblocks support vectormachine, decision trees which performs howeverwith advanced ensemble methods logit leaf model.