A proposed hybrid framework to improve the accuracy of customer churn prediction in telecom industry

Customers are considered one of the most important assets for businesses in numerous dynamic and competitive companies within the marketplace [1].

Competitive market companies, in which customers have numerous choices of service providers, can easily switch a service or even the provider. Such customers are referred to as churned customers [1].

Acquiring a new customer not only costs 5–6 times more than retaining a customer but also requires time to develop customer loyalty to the service provider, subject to the satisfaction of demanded services. While retaining a customer does not involve any additional marketing or other expenses, attention to the resolution of customer concerns is sufficient in most cases.

Furthermore, long-term customers relationship generates more profit because other competitors do not easily attract them, and they can refer to new customers and eventually become less expensive to attend. Thus, a fractional improvement in customer retention can impact the growth and sustainability of telecom businesses [2].

The issue of customer churn or attrition is not unique to the telecom industry. At some point in time, every company in every industry faces the problem of losing customers to competitors. This is usually a source of large financial losses for companies, as it is considered easier and much cheaper to keep an existing customer than to attract new ones. Several studies and surveys support this finding. Van den Poel and Larivière [3] confirmed this in their study of the importance of the economic value of customer retention in the context of a European financial services company [4].

There are different types of artificial intelligence. Machine learning is an application of artificial intelligence, which is used to predict churners and non-churners; Kaličanin et al. in [5] demonstrated that decision makers need to implement some AI applications. Machine learning (ML) represents the main pillar of artificial intelligence and consists of diverse mathematical models like statistics, probabilistic, and neural networks. These models are applied to large datasets to identify data patterns, learn, or predict output values.

Luckert et al. in [6] defined ML algorithms as supervised, unsupervised, semi-supervised, and reinforcement learning. Supervised learning: In this, some data is already tagged with the correct answer, and the machine is trained using this labelled data. It generates a function that predicts the output based on the input observations. Unsupervised learning: This uses information that is neither classified nor labelled and allows the algorithm to act on that information without guidance. Here, the machine is forced to train from an unlabeled dataset and then differentiate it based on certain characters. In [7], semi-supervised learning was defined as training on a combination of labelled and unlabeled data. This learning method combines a small amount of labelled data with a large amount of unlabeled data during training. Reinforcement learning: learning happens in the environment, and this learning employs rewarding desired behaviors and punishing undesirable ones.

Machine learning (ML) has shown outstanding results in a variety of applications, including speech recognition [8], computer vision [9], medical diagnostics [10], and telecom [11]. Among the many approaches developed in the literature for predicting customer churn, supervised machine learning (ML) techniques are the most widely investigated. Supervised ML involves the development of models that can be learned from labelled data. ML includes a wide range of algorithms, such as decision trees, k-nearest neighbors, linear regression, naive bayes, neural networks, support vector machines (SVM), and genetic programming [12].

Ensemble learning is one of the most popular methods that can be used to improve classification accuracy. These methods are a class of highly successful machine-learning algorithms that combine several models to obtain an ensemble that is supposedly more accurate than its individual members [13, 14]. Widely used ensemble classification methods include bagging [15], AdaBoost [16], random forest [17], random subspace [18], and gradient boosting [19, 20]. Because ensemble classification methods have advantages in terms of accuracy, stability, and generalization, they are widely adopted to solve various problems such as multiple-label learning and imbalance learning.

As far as preprocessing is concerned, in our study, the data preprocessing can often have a significant impact on generalization performance of a supervised ML algorithm. The elimination of noisy instances is one of the most difficult problems in inductive ML [21, 22].

In [23], data preprocessing methods that include data cleaning, integration, transformation, and reduction were proposed. In [24], the authors survey the most popular preprocessing steps required for environmental data analysis and present a proposal to systematize them. As in [24], the authors represented Fig. 1, which shows that it yields significant advantages when optimizing the competitive algorithms and tools used to enhance data quality. The aim of this study is to expand on the results of previous studies in which, in the preprocessing layer as well as the classification layer, which still don’t satisfy telecommunication providers as a result, it spotted light from both layers. To predict churn and improve accuracy, a hybrid framework based on the chosen ensemble learning classifier is used to predict churn. Preprocessing is essential for obtaining good-quality results and useful new knowledge from them.

The purpose of this study is to minimize the misclassification of churners and non-churners. A hybrid framework based on the integration of ensemble machine learning and hybrid resampling techniques with the selection of the best techniques for improving data quality is developed to predict customer churn. Previous research on data preparation and preprocessing within a machine learning cycle is argued. Also, it focuses on the limitations of accuracy and performance trade-offs to appropriately target churners, facilitating decision-makers to offer suitable retention plans. Ensemble learning is one of the most popular methods that can be used to reduce the variance in models and improve classification accuracy. The success of XGBoost is primarily due to its scalability in all applications and its speedy learning due to distribution and parallel computation, which allow for quicker model exploration. The SMOTE-ENN method is one of the well-known methods that combines SMOTE as an over-sampling technique and ENN as an under-sampling technique to improve the results. The oversampling led to overfitting the Churner class, resulting in incredibly low precision. Undersampling increased the recall but lowered the precision compared to the baseline score. They suggested opening a research gate for combining techniques. In this study, combining SMOTE and ENN can help determine the benefits and drawbacks of both techniques. Standardization, feature selection, feature reduction, missing value imputation, SMOT-ENN, and XGBoost ensemble learning classifiers are applied in this study. These techniques are applied to improve the accuracy of customer churn prediction.

This study compares pre- and post-data balancing performances for the classifiers, identifies the most important variables that affect customer turnover, and focuses on the speed-accuracy trade-off in hybrid classifiers. It reports an analysis of three datasets with two different experiments using ROC and AUC curves. It has been shown that the result of experiment (a) with the third dataset achieved the highest accuracy of 99.92%, the highest f-score, and the best AUC of 99%, in which hybrid resampling using the SMOTE-ENN technique and STD scaling were performed. While the lowest accuracy level was in experiment (b) in which the resampling is ignored.

The rest of this study is structured as follows: Section "Related Work" presents the previous studies to predict Churn. Section "A proposed hybrid framework for customer churn prediction" presents the main stages of a developing system to predict Churn based on the Proposed Framework. Section "Experiment Results and Discussion" presents experimental results and discussion. Finally, the conclusions are presented in Section "Conclusion".

In this section, the data preprocessing, key concepts, and criteria of its applications will be discussed. It’s worth mentioning the impact of data preprocessing methods, as well as the classification layer, on model performance and the ability to derive best practices within churn management. In addition to that, the study of data mining and machine learning applications of classification within the related domains of the telecom sector will be discussed.

In [25], a study presented an approach to the classification problem with non-discrimination constraints. This classification is based on preprocessing the training dataset. The proposed solution is massaging, reweighing, and sampling. However, they first theoretically studied the trade-off between discrimination and accuracy in a general setting. Similarly, they pointed out the importance of preprocessing and its impact on accuracy.

A recent study in [26] developed a methodological guideline for studies focused on using machine learning. It included a detailed checklist of the eight items and demonstrated detailed data preparation as a list of tasks. These guidelines would improve the quality of research methods.

The study in [7] investigated the influence of data preprocessing and the effects of over- and under-processing. It analyzed a comparison of the popular data preprocessing techniques and their effects on different data classification algorithms. It has been observed that when the dataset was preprocessed using standardization, the best accuracy of 98.24% was obtained. The paper has shown strong over- and under-processing effects when a is overly preprocessed, and the performance of the algorithm decreases.

Regarding the key concepts of preprocessing, as in the prior study [27], the authors suggested a research framework for customer churn prediction. They performed experiments on independent feature selection methods and independent prediction algorithms. They experimentally compared them to construct a two-phase optimization algorithm, although they selected the best match between the feature selection method and the prediction algorithms. This study examined the significance of feature selection in increasing accuracy levels.

In [28], the aim was to predict customer churn using big data. Statistical analyses and J48 decision trees were constructed for the three different datasets. In the preprocessing stage, to provide discriminating features to the classifier, Fisher’s ratio and F-score feature extraction methods were used. This study showed that data preprocessing and normalization are indispensable for better comparability of usage trends between months. They focus on the quality of data, which leads to the quality of decisions.

In a recent study by Zhou and Ooka [29], the paper addressed the impact of preprocessing in terms of time concerns. They found that general preprocessing methods, such as standardization, normalization, and other methods such as a number, proportion, and nondimensionalization for input and output, can result in similar performance on a constructed NN model. This study strongly recommends preprocessing for the output, particularly for situations with variables having different orders of magnitude in the output.

The return to study in [14] has been identified as a major contributing factor for prediction accuracy. They used the AdaBoost algorithm, which is a boosting ensemble method. In this study, the hybrid approach consisted of a support vector machine (SVM) classifier with feature selection. The experimental results showed that the proposed hybrid approach achieved better performance than the base classifier SVM alone. An accuracy of 99.24% was obtained using the hybrid approach. However, the study might have been comprehensive if the authors had considered data resampling techniques to improve accuracy within their model without overfitting.

Similarly, Lalwani in [30] argued that data preprocessing is an important phase. They highlighted the role of feature selection to improve classification accuracy and applied a gravitational search algorithm to perform feature selection. This study proved that the XGBOOST and CatBoost classifiers have significant results in terms of accuracy rates.

Considering the findings reported by Rocha et al. [31], they proposed a model based on a rough set theory using four different rule generation algorithms. This study investigated the performance of two oversampling techniques, SMOTE and MTDF. In addition, they used mRMR feature extraction to improve the classification performance and reduce the computational cost and complexity by eliminating unnecessary features from the dataset. The experiments showed that the genetic algorithm yields the best performance using MTDF for oversampling to manage the imbalanced class problem. Overall, these studies highlight the need to combine resampling techniques to achieve improved results.

For instance, returning to the study in [24], the authors proposed a general preprocessing methodology. A systematic study of preprocessing tasks was reported by the authors and believed to be linked to this review for systematic dealing with preprocessing and data preparation phases, which might be an interesting concern in applying research. As they pointed out, while trying to balance data by oversampling the minority classes, under sampling the majority class, or a combination of both, and using ensemble methods, policies are also used in some cases to gain robustness.

In [32], resampling techniques were utilized to solve class imbalance issues. It has been reported that oversampling led to overfitting the Churner class, resulting in incredibly low precision. Under sampling increased the recall but lowered the precision compared to the baseline score. They suggested opening a research gate for combining techniques.

Earlier research [33] demonstrated that when they investigated the capabilities of preprocessing techniques in an anomaly detection environment. The authors focused primarily on the issue of missing data insertion and data normalization, and the conclusion of the results is quite surprising. The preprocessing techniques for the replacement of missing values are insignificant. Moreover, scaling the dataset can have a considerable impact on the detection capabilities of the anomaly detection system. The paper is limited to their case study or might be the dataset, as these preprocessing techniques enhanced the model performance in other cases.

Another study [34] investigated the impact of combining the resampling techniques of oversampling and under sampling to solve an imbalanced dataset and improve prediction accuracy. It presented a hybrid approach to enhance prediction accuracy. The proposed study was conducted in two phases: (1). Novel hybrid approach: SMOTE oversampling and borderline under sampling (SOS-BUS) (2). Classifier ensemble formation. Experiments performed with different resampling methods were applied to the churn dataset and showed that the hybrid model outperformed the other reference techniques in terms of the area under the ROC curve AUC.

Another study [35] introduced a hybrid model that improved the results of predictions based on an SVM. They used a preprocessing method that combines oversampling and under sampling to achieve more accurate results.

In [36], the authors proposed the accuracy of churn prediction using a deep learning model with machine-learning algorithms. The random forest model has given better accuracy, which is approximately 85%. The study demonstrated the impact of hyperparameter optimization on performance.

The authors in this study apply a three-stage technique for customer churn prediction. In the first stage, the feature selection technique has been used to choose the most relevant features by removing the redundancy and maximizing the significance. In the second stage, the ripple-down rule has been used to build a knowledge-based system to gather knowledge about the customer’s behavior. In the final stage, the knowledge acquisition is evaluated by the simulated expert technique. The proposed approach is applied to the churn dataset. It outperformed well for churn prediction in the telecommunications industry, with an accuracy of 95.169% [37].

This study addressed the problem of cross-company churn prediction (CCCP), where the company that called a target doesn’t have enough data and should depend on data from another company that called a source to successfully predict customer churn. Until now, there has been no effective method to transfer data in CCCP. It introduced data transformation methods for CCCP using z-score, log, rank, and box-cox. It measured the effect of these methods on CCCP using different classifiers like k-nearest neighbor, naive bayes, single rule induction, gradient boosted tree, and deep learner neural networks for predicting customer churn in telecommunication companies. The experiments were applied to customer churn datasets in telecommunication companies. The results confirmed that the proposed data transformation methods based on log, rank, and box-cox improved the performance of CCCP and achieved better results using naive bayes [38‐40].

The authors confirmed the importance of applying the proposed approach to evaluate the classifier’s reliance on the used data and estimate its accuracy. The certainty estimation of the classifier is done by calculating the correlation of all the dataset's features using the normal distribution’s six-sigma rules. The approach is applied to extract the classifier’s certainty level of decision. It classified the consumers into different groups according to the lower and upper zones. Then the evaluation of the certainty of the classifier is measured before classifying the consumer churn and non-churn [41].

The authors confirmed a lack of effective, rule-based to predict customer churn in telecom companies. They proposed a technique-based intelligent rule to make accurate decisions to classify churn from non-churn customers and identify the customers who are likely to churn or may churn soon [42].

The authors highlighted the important impact of efficient manipulation of a high-dimensional dataset. They said that scientific papers applied feature reduction to reduce processing time and enhance accuracy. However, the authors confirmed that feature reduction causes the loss of important information. They confirmed that assigning weights to the features and avoiding information loss by domain experts is effective, but it is expensive and requires human expertise in the domain. They proposed a novel automatic technique to assign weights without needing domain experts. The results have shown that the proposed technique achieved an accuracy of 89.1% and outperformed feature reduction [43].

As mentioned above and in Table 1, while numerous factors could impact the accuracy level of those churners during the prediction model, most fields of study primarily concentrate on classification algorithms that increase churn prediction accuracy levels. Previous studies often utilized algorithms like decision trees, ANN, logistic regression, Naïve Bayes, and K-nearest neighbor. Limited studies are focused on data preparation, which accounts for more than half of the learning models in the total data discovery process and requires more time and effort. Many studies have emphasized methodical data preparation. In practice, applying the preprocessing processes sequentially is unnecessary, which leads to their implementation as a pipeline of tasks. This is based on the nature of the dataset used. The analysis and integration of appropriate preprocessing techniques and capabilities received insufficient attention from the research community. Advanced ensemble learning models like XGBoost provide excellent predictions on classification issues. Only a limited number of studies have used XGBoost to predict customer churn. The datasets deployed to estimate customer churn are frequently unbalanced, containing many non-churn instances and very few churn instances. To balance the data, previous work has mostly used the Synthetic Minority Oversampling Technique (SMOTE). Hybrid resampling techniques like SMOTE-ENN have been suggested as unique and efficient techniques. Few studies have used this hybrid approach to predict customer churn. The achievement of applying a hybrid framework that combines the benefits of selecting the proper techniques for improving data quality, resampling methods, and ensemble learning hasn’t been studied before. As a result, this study intends to provide a novel contribution to research in customer churn by constructing a prediction model combining suitable techniques for data preprocessing with hybrid resampling methods and ensemble learning algorithms. It compares the model's prediction performance with earlier studies. The publicly accessible telecom dataset is subjected to the XGBOOST classifier. To produce more accurate results, efficient techniques for improving data quality and a hybrid preprocessing technique that combines oversampling and undersampling techniques have been employed. The model used a stratified k-fold to create a balance between training and test data and improve classification accuracy. Also, it covers a variety of input data by verifying the model's performance on several folds.

In the next section, an overview of the proposed framework has been presented for accurate prediction. In addition, the results of the proposed hybrid framework are compared with those of previous studies using the same datasets, and the results of this comparison are presented.

The hybrid framework for customer churn prediction addresses the key concepts for the accuracy of customer churn prediction. The proposed framework introduces a solution over three layers: the data, algorithm, and evaluation layers, as shown in Fig. 2.

This study proposes a hybrid framework for customer churn prediction that addresses the key concepts for churn prediction accuracy. It presents a comparative analysis of two experiments with three datasets. The proposed framework introduces a solution over three layers: data, algorithm, and evaluation. In phase one, the data layer represents the data preprocessing phase. It conducted two main tasks: data preparation and feature engineering. In phase two, the algorithm layer integrates two resampling techniques—SMOTE-ENN, a preprocessing step to reduce the negative effects caused by a class imbalance—and applied the XGBOOST classifier as an ensemble learning algorithm. In phase three, the evaluation layer applies four classification evaluation methods: accuracy score, precision, recall, and F-score. They are used to validate the results, and the results of cross-validation and the testing data are registered. This study offeres results that outperformed the findings of previous works in this field with the same datasets mentioned above in Table 9. Hence, our proposed hybrid framework shows the highest performance and accuracy.