(IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 12, No. 11, 2021 277 | Page www.ijacsa.thesai.org Improving Customer Churn Classification with Ensemble Stacking Method Mohd Khalid Awang, Mokhairi Makhtar, Norlina Udin, Nur Farraliza Mansor Faculty of Informatics and Computing, Universiti Sultan Zainal Abidin, 22000 Tembila, Terengganu, Malaysia Abstract—Due to the high cost of acquiring new customers, accurate customer churn classification is critical in any company. The telecommunications industry has employed single classifiers to classify customer churn; however, the classification accuracy remains low. Nevertheless, combining several classifiers' decisions improves classification accuracy. This article attempts to enhance ensemble integration via stack generalisation. This paper proposed a stacking ensemble based on six different learning algorithms as the base-classifiers and tested on five different meta-model classifiers. We compared the performance of the proposed stacking ensemble model with single classifiers, bagging and boosting ensemble. The performances of the models were evaluated with accuracy, precision, recall and ROC criteria. The findings of the experiments demonstrated that the proposed stacking ensemble model resulted in the improvement of the customer churn classification. Based on the results of the experiments, it indicates that the prediction accuracy, precision, recall and ROC of the proposed stacking ensemble with MLP meta-model outperformed other single classifiers and ensemble methods for the customer churn dataset. Keywords—Stacking ensemble; customer churn prediction; bagging; boosting I. INTRODUCTION The rapid development of wireless telecommunications has altered the course of Malaysia's telecommunications industry [1]. Customers may choose and switch between the packages of various service providers. Churn is a term used to describe the behaviour of customers who switch service providers, and it has become a significant issue for Malaysian network providers. Numerous researchers have attempted to develop various classifiers to predict customer churn, including the decision tree [2], genetic algorithm [3], and regression analysis [4]. However, the conventional approach of using single classifiers for churn prediction is ineffective. It should be improved, as various uncertainty factors such as customer service, network coverage, product quality, packaging prices, and reception quality can all contribute to customer churn [5]. Furthermore, a set of classifier methods referred to as the ensemble method may be used to improve prediction accuracy. The ensemble approach performs better than individual classifiers because of their divergence or independent character. The ensemble technique combines the choices of many classifiers to enhance classification performance [6]. Multi-classifier ensemble techniques, also known as many classifiers, are machine learning algorithms that include training many base classifiers and then aggregating their output to get the highest possible prediction accuracy [7]. Combining the predictions of several classifiers, such as bagging [8], boosting [9], stacking [10] and ensemble selection [11], maybe a practical approach for improving classification performance. The rest of this article is structured as follows: Section 2 discusses the review of related literature, including ensemble methods such as bagging, boosting, and stacking. Section 3 covers the research methodology, including the data set and the proposed ensemble stacking. Section 4 presents the experimental setup and results from the discussion. The conclusion of this research is discussed in Section 5. II. LITERATURE REVIEW A. Predictive Analytics Predictive analytics is the most often used technique of predicting customer turnover in the business world. When it comes to predictive modelling, it is a model that can be used to forecast or estimate the target values of future instances [12]. In the context of this research, it is described as the process of forecasting or identifying consumers who are likely to abandon their current purchases in the near future [13]. Predictive analytics is made up of a variety of techniques such as statistical prediction modelling, machine learning modelling, and data mining that analyse previous information and make predictions about future events or something completely new and unknown [14]. Predictive modelling is a technique in which a classifier is usually built based on certain information in order to anticipate the result of a given situation. In accordance with [15], predictive modelling may be divided into four subcategories, as follows: 1) Classification is used when the predicted result is categorical in nature. 2) A regression analysis is used when the prediction results in a numerical value as the result of the analysis. 3) Clustering is the term used to describe the process of grouping a certain collection of items based on their characteristics as a result of the analysis. 4) When the result is the discovery of intriguing connections between data, this is referred to as association rules.