An ensemble method with cost function on churn prediction

• Main Authors: Mohd Khalid, Awang, Mohammad Afendee, Mohamed, Mokhairi, Makhtar
• Format: Conference or Workshop Item
• Language: English; English
• Published: 2019
• Subjects: HG Finance
• Online Access: http://eprints.unisza.edu.my/1874/1/FH03-FIK-20-36862.pdf; http://eprints.unisza.edu.my/1874/2/FH03-FIK-20-36863.pdf; http://eprints.unisza.edu.my/1874/
• Institution: Universiti Sultan Zainal Abidin

Accurate customer churn classification is vital in any business organisation due to the higher cost involved in getting new customers. In telecommunication businesses, companies have used various types of single classifiers to classify customer churn, but the classification accuracy is still relatively low. However, the classification accuracy can be improved by integrating decisions from multiple classifiers through an ensemble method. Despite having the ability to produce higher classification accuracy, the ensemble method tends to produce similar or redundant classifiers. Therefore, this paper aims to achieve higher classification accuracy and at the same time, minimising ensemble classifiers by constructing a new ensemble method based on dimensionality reduction in soft set theory. The combination of ensemble classifier is calculated based on the simple majority voting algorithm. The performance measure used in determining the optimal subset of classifiers is the combination of Accuracy (ACC), True Negative Rate (TNR) and True Positive Rate (TPR). The proposed soft set ensemble methods (SSPN and SSSC) are systematically evaluated using customer churn data set taken from one of the local Telco companies in Malaysia. The selection and combination algorithm (SSSC) has proven its supremacy by producing accuracy (ACC) of 87.0% for local Telco data set and 94.0% for UCI data set, which is better than any other single classifier. This work proved that the proposed soft ensemble method could search for the minimum number of classifiers in the ensemble repository while at the same time improving the classification performance. In conclusion, the proposed soft ensemble method not only reduces the number of members of the ensemble but is also able to produce the highest classification accuracy