A False Negative Cost Minimization Ensemble Methods for Customer Churn Analysis

Abstract

The primary objective of this research is to develop hybrid decision tree induction methods based on the decision tree C4.5 algorithm and ensemble methods, taking into account cost-sensitivity for the purpose of minimizing either misclassification cost, false negative cost or false positive cost. This paper proposed two cost-sensitive learning methods by modifying the model weight of AdaBoost.M1 for churn analysis in the telecommunication industry. Method 1 applies the ratio of false negative cost over true negative cost to make the weight of false negative heavier than the weight of false positive. While Method 2 combines error rate weighting with false negative cost weighting in order to let examples have heavier weight values for future training in the next learning cycle. The proposed methods have been evaluated with a series of experiments to prove its ability to reduce either false negative cost or misclassification costs. Microsoft Azure Machine Learning Telco Customer Churn and IBM Watson Studio Telecommunication Customer Churn datasets, which include the cost value for each instance, are used for the experiments. The proposed Method 1 able to obtain the lowest false negative cost comparing with the original AdaBoost.M1.