Revolutionizing market surveillance: customer relationship management with machine learning

Machine learning in customer relationship management

Machine learning (ML) has indeed recently begun to draw a lot of attention from the theory of CRM since it is capable of dealing with huge amount of customer’s information and making dynamic prediction. As mentioned in the article (Abdullah & Jalil, 2022), the use of ML played a critical role to improve the existing CRM strategies by analyzing customers’ behaviour and their interactions. More and more applied ML models, such as decision trees, random forests, and neural networks are being used to enhance the customer segmentation, churn prediction, marketing employments efforts (Abuhasel, 2023). However, several literatures should have established that various single ML models have applicability in CRM; literature on the use of multiple models to boost the accuracy and the resilience of the CRM systems is rather scarce. This gap requires the use of a more complex study that integrates the effectiveness of other models which is the focus of this research.

Applications of ensemble learning in CRM

Recent studies have attempted to observe the ensemble learning, a method in which multiple ML models are fed to increase the overall prediction capability of the system. The article (Alhaqui, Elkhechafi & Elkhadimi, 2022) presented an intelligent CRM system employed with the theory of fuzzy clustering for constructing customer segments. The current system improved the accuracy of the analysis of the customers compared to the system that uses single models. Likewise, in the article (Almahairah, 2023) a prototype implementing the Recency, Frequency, and Monetary (RFM) model has been developed using machine learning to categorize customer so as to improve the chances of rendering a more effective marketing strategy. However, these studies primarily focus on single tasks like segmentation, leaving a gap in research addressing how ensemble learning can improve multiple CRM functions, such as churn prediction and customer retention strategies.

In the context of churn prediction, this study (Alojail & Bhatia, 2020) applied machine learning to predict customer churn in the telecom sector using single models like random forest and support vector machine. Their findings show promise but highlight the limitations of using a single model to handle complex and dynamic customer data. Our research builds on this by developing an ensemble learning-based model that integrates several ML techniques to overcome the limitations of single models, offering a more robust churn prediction system.

Despite the growing body of literature, several gaps remain unaddressed. Most studies focus on applying machine learning in isolation, either for customer segmentation, churn prediction, or personalized marketing. Few explore the potential of ensemble learning to simultaneously optimize these multiple CRM functions. In addition, even though there has been a great focus on technical evaluation criteria (accuracy, precision, and recall), much less effort has been directed towards the incorporation of such systems into decision support systems which are extremely useful in the actual application of intelligent systems.

Ethical concerns and bias in machine learning for CRM

Yet another research area that has not been well explored in the literature is the aspect of ethics including, data privacy and algorithmic bias. To this end, this study (Asha et al., 2023) provided a literature-based analysis of the ethical issues surrounding the use of ML for CRM, particularly with regard to the lack of data openness and the problem of biased models. However, their work did provide no real applicable strategy of managing these risks. More recently, the article (Batra & Rehman, 2019) illustrated how machine learning algorithms can transport unfairness to certain customer groups, when using CRM. Many such ethical considerations are still missing in the ensemble learning, which could have new implications. To ease these concerns, this research aims at designing a better and clear approach of integrating the models such that the ensemble learning system is efficient yet socially sensitive. The comparative table of previous study is indicated in Table 1. This research aims to fill these gaps by proposing the SmartSurveil CRM model, an ensemble machine learning approach integrated within a DSS to ensure ethical and effective churn prediction in CRM.

Research gaps

In summary, the current state of the literature reveals several gaps:

• Lack of ensemble learning integration: While individual machine learning models have been extensively studied, the integration of multiple models for improving CRM system performance, particularly in churn prediction, remains underexplored.

• Practical application in decision support systems (DSS): Although ML models have been shown to enhance CRM, few studies explore how these models can be effectively integrated into real-time DSS for customer retention.

• Ethical considerations: There is limited research on how to address data privacy, bias, and model transparency in CRM systems using machine learning, particularly when multiple models are combined.

• This study aims to address these gaps by developing the SmartSurveil CRM model, an ensemble-based machine learning system integrated into a DSS, offering robust customer churn prediction while addressing the ethical challenges associated with machine learning in CRM.

Problem formulation 1: data processing and analysis

The problem of data processing and analysis in market surveillance and CRM can be formulated as follows. Given a dataset $\mathrm{D}=\{\left({\mathbf{x}}_{\mathrm{i}},{\mathrm{y}}_{\mathrm{i}}\right){\}}_{\mathrm{i}=1}^{\mathrm{N}}$, where ${\mathbf{x}}_{\mathrm{i}}\in {\mathbb{R}}^{\mathrm{d}}$ represents the features and ${\mathrm{y}}_{\mathrm{i}}$ represents the corresponding labels or outcomes, the objective is to develop efficient techniques for processing and analyzing this data to support real-time decision-making and insights generation.

(1) $$\hat{\mathrm{\theta }}=\mathrm{a}\mathrm{r}\mathrm{g}\underset{\mathrm{\theta }}{\mathrm{m}\mathrm{i}\mathrm{n}}{\displaystyle \frac{1}{\mathrm{N}}\sum _{\mathrm{i}=1}^{\mathrm{N}}\mathcal{L}\left({\mathrm{y}}_{\mathrm{i}},\mathrm{f}\left({\mathbf{x}}_{\mathrm{i}};\mathrm{\theta }\right)\right)+\lambda \mathrm{R}\left(\mathrm{\theta }\right)}$$where $\mathrm{f}\left(\mathbf{x};\mathrm{\theta }\right)$ is a model parameterized by $\mathrm{\theta }$, $\mathcal{L}$ is the loss function, $\mathrm{R}$ is the regularization term, and $\lambda$ controls the trade-off between the data fitting and regularization.

(2) $$\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{m}\mathrm{i}\mathrm{z}\mathrm{e}\mathrm{J}\left(\mathrm{\theta }\right)={\displaystyle \frac{1}{\mathrm{N}}\sum _{\mathrm{i}=1}^{\mathrm{N}}\mathcal{L}\left({\mathrm{y}}_{\mathrm{i}},\mathrm{f}\left({\mathbf{x}}_{\mathrm{i}};\mathrm{\theta }\right)\right)+\lambda \mathrm{R}\left(\mathrm{\theta }\right).}$$

The objective is to minimize the loss function $\mathcal{L}$ and the regularization term $\mathrm{R}$ to find the optimal parameters $\hat{\mathrm{\theta }}$ that best fit the data while avoiding overfitting.

• $\mathrm{D}$: Dataset containing $\mathrm{N}$ samples.

• ${\mathbf{x}}_{\mathrm{i}}$: Feature vector for the $\mathrm{i}$th sample.

• ${\mathrm{y}}_{\mathrm{i}}$: Label or outcome for the $\mathrm{i}$th sample.

• $\mathrm{\theta }$: Model parameters.

• $\mathcal{L}$: Loss function measuring the discrepancy between predicted and true labels.

• $\mathrm{R}$: Regularization term to prevent overfitting.

• $\lambda$: Regularization parameter.

The problem of data processing and analysis aims to develop techniques to handle the variety, volume, and velocity of data in market surveillance and CRM systems. The objective is to process and analyze this data efficiently to support real-time decision-making and insights generation. Mathematically, this problem involves finding the optimal parameters $\hat{\mathrm{\theta }}$ of a model that minimize a loss function $\mathcal{L}$ and a regularization term $\mathrm{R}$. These parameters are learned from the dataset $\mathrm{D}$, which consists of feature vectors ${\mathbf{x}}_{\mathrm{i}}$ and corresponding labels ${\mathrm{y}}_{\mathrm{i}}$. The objective function $\mathrm{J}\left(\mathrm{\theta }\right)$ represents the trade-off between fitting the data and regularization, controlled by the regularization parameter $\lambda$.

Problem formulation 2: real-time adaptability

The problem of real-time adaptability in CRM systems can be formulated as follows. Given a set of customer interactions $\mathcal{J}=\{\left({\mathbf{x}}_{\mathrm{i}},{\mathrm{y}}_{\mathrm{i}}\right){\}}_{\mathrm{i}=1}^{\mathrm{N}}$, where ${\mathbf{x}}_{\mathrm{i}}$ represents the features of the interaction and ${\mathrm{y}}_{\mathrm{i}}$ represents the outcome (e.g., customer satisfaction score), the objective is to design dynamic and adaptive algorithms that can continuously learn from incoming data, adjust to changing market dynamics, and provide personalized customer interactions in real-time, while ensuring compliance with regulatory requirements such as the Basel equation for risk assessment.

(3) $$\hat{\mathrm{f}}\left(\mathbf{x}\right)=\mathrm{a}\mathrm{r}\mathrm{g}\underset{\mathrm{f}\in \mathcal{F}}{\mathrm{m}\mathrm{a}\mathrm{x}}{\mathbb{E}}_{\mathcal{D}}\left[\mathrm{L}\left(\mathrm{f}\left(\mathbf{x}\right),\mathrm{y}\right)\right]-\lambda \mathrm{R}\left(\mathrm{f}\right)$$where $\hat{\mathrm{f}}\left(\mathbf{x}\right)$ is the learned model function, $\mathcal{F}$ is the set of possible model functions, $\mathcal{D}$ is the distribution of the data, $\mathrm{L}$ is the loss function measuring the discrepancy between predicted and true outcomes, and $\mathrm{R}$ is the regularization term accounting for compliance with the Basel equation.

(4) $$\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{m}\mathrm{i}\mathrm{z}\mathrm{e}\mathrm{J}\left(\mathrm{f}\right)={\mathbb{E}}_{\mathcal{D}}\left[\mathrm{L}\left(\mathrm{f}\left(\mathbf{x}\right),\mathrm{y}\right)\right]-\lambda \mathrm{R}\left(\mathrm{f}\right).$$

The objective is to minimize the expected loss $\mathrm{J}\left(\mathrm{f}\right)$ by learning a model function $\mathrm{f}$ that accurately predicts outcomes $\mathrm{y}$ based on customer interactions $\mathbf{x}$, while ensuring compliance with regulatory requirements such as the Basel equation.

• $\mathcal{J}$: Set of customer interactions.

• ${\mathbf{x}}_{\mathrm{i}}$: Feature vector for the $\mathrm{i}$th interaction.

• ${\mathrm{y}}_{\mathrm{i}}$: Outcome for the $\mathrm{i}$th interaction.

• $\hat{\mathrm{f}}$: Learned model function.

• $\mathcal{F}$: Set of possible model functions.

• $\mathcal{D}$: Distribution of the data.

• $\mathrm{L}$: Loss function measuring the discrepancy between predicted and true outcomes.

• $\mathrm{R}$: Regularization term accounting for compliance with regulatory requirements.

• $\lambda$: Regularization parameter.

Dataset description

The dataset used for this study consists of customer data from a telecom company. The dataset includes various features related to customer demographics, behavior, and interactions with the company. The target variable for this study is customer churn, which indicates whether a customer has left the company or not. The dataset contains the following features. The telecom dataset used in this study was obtained from an internal customer database of a telecom company, which included various customer attributes such as demographics, behavior, and interaction records. Before preprocessing, basic filtering was applied to remove duplicate and incomplete records, ensuring the dataset’s integrity. The final dataset comprised X number of records (you can specify the exact size). For model training and evaluation, 70% of the data was allocated to training, while 30% was reserved for testing. Random Forest (RF) was chosen for the proposed model due to its robustness in handling large datasets and its ability to prevent overfitting by combining the results of multiple decision trees. RF is also known for its high accuracy in classification tasks and its capability to handle both categorical and continuous variables, making it suitable for churn prediction in the telecom sector. The Table 2 shows the features description of the dataset. The telecom internal database used in this study included anonymized customer data, removing any identifiable personal information to protect individual privacy. This anonymization process adhered to data protection regulations, including GDPR guidelines where applicable, ensuring that all data was de-identified before analysis. Additionally, the research process involved stringent access controls, allowing only authorized personnel to handle the data, further minimizing risks associated with unauthorized use or exposure. In terms of algorithmic bias, steps were taken to assess and mitigate potential biases in the model, ensuring fair and unbiased predictions across diverse customer groups. Our approach underscores a commitment to ethical data usage, prioritizing customer privacy, data security, and regulatory compliance throughout the study.

The correlation heatmap of the CRM features is presented in Fig. 4. The closer that two features are connected and the higher the absolute value of coefficients, the stronger their relationship is according to this heatmap. The darkness of the color signifies the real value or the degree of correlation coefficients as derived from the data. This type of visualization is useful to determine the degree of association of all the input features with the target feature and between themselves, when it comes to performing feature selection/engineering.

The overall idea behind this is given in Fig. 5—this is the 3D scatter plot that represents the customer segmentation according to age, annual income, and numerous spending score. Customers are segmented regarding three characteristics, and the color panels used make it easy to understand the relationship between these three variables and possible clusters. It assists to describe patterns of behaviour of the customers and their distribution and is specific to the areas of marketing and selling.

Figure 6 below shows the count plot for gender and churn. The plots depict the trends in the proportion of male and female customers identified in the dataset; moreover, churned customers and customers who did not churn are represented as well. A critical aspect of customer data analysis is understanding customer attrition and its distribution either by gender in this case; thus, the visualizations below depicting the percentage of churn among the identified genders:

Histograms for the different CRM features are shown in Fig. 7. These histograms present different features like the age, annual income, the spend score and so much more. This information is significant for activities, such as normalization and can help to provide a view on customers’ behavior.

The structural similarity of Fig. 8 shows the box plots of the specified features related to the churn variable. The following box plots offer a quick flag of distribution and variability of the features of churned customers compared to non-churned customers. They assist in assessment of variance of feature values between these two groups and can therefore be useful in feature subset and model derivation.

To better understand the metrics of CRM features, the distribution plots are shown in Fig. 9. Using such plots, one is able to assess the distribution of the different feature usually characterized by mean, variance, and skewness or kurtosis. Pictorially the distribution plots helps on checking for outlying values and on interpretation of the nature of the data.

Proposed model: SmartSurveil CRM

The proposed model, SmartSurveil CRM, utilizes the Random Forest algorithm to predict customer churn. Random Forest is an ensemble learning method that operates by constructing multiple decision trees during training and outputting the mode of the classes for classification tasks. Below, we provide a detailed mathematical formulation of the Random Forest algorithm as applied in SmartSurveil CRM.

Figure 10 presents the internal structure of the proposed SmartSurveil CRM model that is intended for the efficient churn prediction in the telecommunications market. This architecture uses the ensemble learning, which involves the use of other sub systems of machine learning in order to make the predictions more accurate and more secure. The key components of the architecture are as follows:

1. Input Data: The detailed information of the customers that may be directly obtained from customer, their characteristics, their interactions and their activities.

2. Feature Selection: It involves the identification of which features within the input data is most important and has high correlation with the target variable—the customer churn. Therefore, it is essential to select relevant features, which would help the development of the model and simplify calculations.

3. Random Forest Model: One of the single learning models that are used together to form the ensemble. Another algorithm that comes under the classification header is the Random Forest algorithm which builds multiple decision trees and returns the mode of the class in case of classification.

4. Gradient Boosting Model: It is another base model in the ensemble. Gradient Boosting develops models on the successive cycle, where each attempt to minimize the residual mistakes from the prior models. It is beneficial in eliminating bias and high variance in model predictions.

5. Support Vector Machine (SVM): The third base model in the ensemble is the logistic regression model which is used in this ten-fold cross-validation. SVM is an effective grouping algorithm that identifies the best line that separates instances of various categories to the highest degree.

6. Ensemble Aggregation: This component collates the output from the models which include random forest, gradient boosting and support vector machine. The aggregation method, which could be another classification technique used was a voting system where by arriving at a final class, the resultant class from the optimization of each model is considered as the final result.

7. Final Prediction: The last stage of the model which helps to target a specific customer as a churning or non-churning customer out of the ensemble of all the models.

8. Hyperparameter Tuning: At each step, hyperparameters of each base model are selected using grid search and cross-validation methods for the best possible performance of the entire model. This tuning allows for all models to be helpful in creating the ensemble.

The strengths of different ML models are incorporated in this architecture to make SmartSurveil CRM is a structurally robust and reliable predictive tool. This tool can than be used to identify customer churn and further guide the decision-making process in the telecom industry.

Random forest algorithm

The second method also known as the Random Forest addresses the issue of over-fitting by stacking many iterated decision trees. The steps involved are as follows:

Bootstrap sampling: It is possible to take an aliquot of n samples at random from the dataset and form a number of sets.

Decision tree construction: For each bootstrap sample grow a decision tree. At every node, all features are randomly selected and the one that provide maximum split according to a measure like Gini index or entropy is chosen.

Aggregation: The predictions from all the decision trees should be combined together in an effort to arrive at the final decision. The mode of the predictions is used for classification tasks.

Mathematical formulation

Hyperparameter tuning

The performance of the SmartSurveil CRM model is optimized by tuning the hyperparameters of each base model in the ensemble. The following table summarizes the hyperparameters and their values for the Random Forest, Gradient Boosting, and Support Vector Machine (SVM) models. The hyperparameter values for the SmartSurveil CRM are shows in Table 3.

Explanation of hyperparameters:

• Number of trees: The number of trees in the random forest.

• Max depth: The maximum depth of the trees.

• Min samples split: The minimum number of samples required to split an internal node.

• Min samples leaf: The minimum number of samples required to be at a leaf node.

• Max features: The number of features to consider when looking for the best split.

• Learning rate: In gradient boosting, the learning rate reduces the impact of each tree for the same learning rate.

• Number of estimators: The number of times boosting has to be performed.

• Loss function: The loss function to be optimized in gradient boosting.

• Kernel: The kernel type to be used in the SVM algorithm.

• C (Regularization): The regularization parameter in SVM.

• Gamma: The kernel coefficient for ‘rbf’, ‘poly’, and ‘sigmoid’ kernels in SVM.

Model evaluation

Several are the metrics used to exhibit the effectiveness of the SmartSurveil CRM model, metrics which undoubtedly reveal the performance of the model. Such measures are accuracy, precision, recall, F1 score and ROC-AUC. In this article, the following equations were used in calculating the different metrics and the table below defines them.

Performance of the proposed model

The performance of the test dataset is then examined in relation to the SmartSurveil CRM model. The criteria chosen are accuracy, precision, recall, F1-score, ROC-AUC. Table 5 presents the outcome measures for the SmartSurveil CRM model.

Figure 11 shows the accuracy comparison for different models used in the study, including SmartSurveil CRM, random forest, gradient boosting, and SVM. The bar chart clearly demonstrates that the SmartSurveil CRM model achieves the highest accuracy among the compared models, indicating its superior performance in correctly predicting customer churn.

We present in the Fig. 12 ROC curves for different models. The ROC curve plots the true positive rate (recall) against the false positive rate for various threshold settings. The area under the ROC curve (AUC) is a measure of the model’s ability to distinguish between classes. The SmartSurveil CRM model’s ROC curve shows a higher AUC compared to other models, indicating better overall performance.

Figure 13 illustrates the learning curves for different models. Learning curves show the training and validation performance as a function of the training set size. These curves help in diagnosing whether a model is suffering from bias (underfitting) or variance (overfitting). The SmartSurveil CRM model’s learning curves indicate a good balance between bias and variance, showing that it generalizes well to unseen data.

Figure 14 presents the levels of confusion between models. On the same note, the confusion matrix is preferred more than the accuracy because it offers a clear distinction of the actual positive and actual negative data by displaying the true positives, true negatives, false positives, and false negative values. The depiction of confusion matrix of the SmartSurveil CRM model reveals more of true positive and true negative values with relatively lower false positive and false negative values which proves the efficiency of the model in providing precise churn rate of the customers. The implication of these findings is a proof of the suitability of the SmartSurveil CRM model in predicting customer churn as presented by the Accuracy of 94%, Precision of 96%, Recall of 91% and F1-score of 94% in addition to the ROC-AUC value of 0.95. Below, these metrics indicate that the predicted churned and non-churned customers are clearly defined and achievable by the model:

Performance comparison of all models

To evaluate the robustness of the SmartSurveil CRM model, we compare its performance with other baseline models including random forest, gradient boosting, and SVM. The comparison metrics include the same evaluation metrics used for the proposed model. Table 6 presents the performance metrics for all models.

The proposed SmartSurveil CRM model has higher recall, precision, F1 score and best F score than the baseline models. A comparison of the results shows that our proposed SmartSurveil CRM achieves notably higher accuracy, precision, recall, F1-score, and ROC-AUC values, which proves the higher predictive ability of the algorithm with regard to customer churn. The approach applied in SmartSurveil CRM involves multiple models, and this makes the system to be robust and accurate since errors associated with one model may be offset by gains in another or by another independent model. Unlike random forest, gradient boosting and support vector machine, SmartSurveil CRM proved to be more effective in the classification of positives as shown in the ROC curves in Fig. 12. The specificity, also known as the false positive rate, is the proportion of negatives that are incorrectly classified as positive; the sensitivity, also known as the true positive rate or recall, is the proportion of positives that are correctly classified as such across different threshold settings.

Key observations

1. Area Under the Curve (AUC): The area under the ROC curve (AUC) is a single scalar value that summarizes the overall ability of the model to discriminate between positive and negative classes. The AUC value ranges from 0 to 1, with higher values indicating better performance. In Fig. 12, the AUC values for each model are:

• SmartSurveil CRM: 0.85

• Random Forest: 0.80

• Gradient Boosting: 0.74

• SVM: 0.66

The SmartSurveil CRM model has the highest AUC value, indicating superior overall performance in distinguishing between churned and non-churned customers.

2. True positive rate (recall) vs. false positive rate: The ROC curve shows how the true positive rate (sensitivity or recall) varies with the false positive rate. A model with a higher true positive rate for a given false positive rate is considered better. In Fig. 12, the SmartSurveil CRM model consistently achieves a higher true positive rate compared to the other models across most of the range of false positive rates.

3. Model Comparison: The ROC curves of the models can be compared to evaluate their relative performance. The curve that is closer to the top-left corner of the plot indicates better performance. In this case, the ROC curve of the SmartSurveil CRM model is closest to the top-left corner, followed by random forest, gradient boosting, and support vector machine. This also suggests that the proposed SmartSurveil CRM model is better than the other models when it comes to the sensitivity and specificities.

The analysis of the curves and their integrated performance measure, AUC, shows that SmartSurveil CRM model provides the highest accuracy in predicting customer churn. The latter also has higher AUC value and better position of ROC curve, which directly points out that it has higher ability to classify churned and non-churned customers with less false positives and false negatives. This enhanced performance affirms both the stability and effectiveness of the SmartSurveil CRM model for real use in customer churn prediction.

Model interpretation

The interpretation of the SmartSurveil CRM model involves understanding the factors that contribute to customer churn and how the model’s predictions can be utilized to make informed business decisions. By analyzing feature importance and model predictions, we can gain insights into customer behavior and the underlying reasons for churn.

Feature importance

Feature importance provides insights into which features have the most significant impact on the model’s predictions. Table 7 lists the top features contributing to customer churn as identified by the SmartSurveil CRM model.

The results in Table 7 indicate that the most critical features influencing customer churn are “Days since Last Purchase”, “Customer Satisfaction”, and “Complaint Count”. These insights can guide targeted interventions to reduce churn.

Decision support system

Decision support system links the findings of the SmartSurveil CRM model with the best strategies for business management. The DSS uses model predictions and feature importance to point out customers that are most likely to churn and recommend appropriate ways to retain them.

Components of the DSS:

• Customer churn prediction: Estimates how likely any given customer is likely to churn.

• Feature analysis: Finds out the features responsible for churn prediction.

• Retention strategy suggestions: It presents specific suggestions to customers and markets as well as feature based conclusion.

Business strategy adjustment

Therefore, using all the information that appears in the PES, managers can make varied decisions that serve to optimize customer retention. Table 8 below shows particular procedures that can be applied based on various customer segments defined by the model.

Implementation of strategies

• Personalized loyalty programs: Offer loyalty points, discounts, and rewards based on customer purchase history and preferences.

• Complaint resolution: Develop a robust complaint management system to handle and resolve customer issues efficiently.

• Targeted marketing: Use targeted marketing campaigns to re-engage customers who show signs of reduced activity.

• Premium services: Offer premium services such as dedicated account managers and exclusive deals to high-value customers.

The SmartSurveil CRM model has a decision support system that offers useful information on which customers their business can possibly lose and then tailors the business strategies for handling different customers. Taking advantage of these and other detailed recommendations in the feature analysis, businesses can increase their positive customer satisfaction, minimize churn rates and ultimately benefit their top and/or bottom lines. The above article also proves the efficiency of using machine learning models in combination with business strategies for improving the customer relationship management.

The authors (Abdullah & Jalil, 2022; Abuhasel 2023) used fuzzy clustering and decision tree based CRM, and Alhaqui, Elkhechafi & Elkhadimi (2022) have put their emphasis on the Random Forest and Support Vector Machine models for churn prediction. Almahairah (2023) employed AI-based fuzzy clustering; the study was useful in customer segmentation. Customer satisfaction sentiment analysis was initiated by Umamaheshwari, Harikumar & Allinjoe (2021), and it focused only on the customer feedback.

To address this problem, our study proposed an ensemble learning model of Random Forest, Gradient Boosting, and Support Vector Machine with providing better churn prediction and retention policies. In all previous studies, customers’ datasets from various industries including telecom and e-commerce were employed but many of these were restricted to only a few applications including the segmentation and feedback.

In our study, telecom data set is used for churn prediction with addition attributes such as demographics and interaction records for broader CRM implementation. Abdullah & Jalil (2022) and Almahairah (2023) got a better customer segmentation while Alhaqui, Elkhechafi & Elkhadimi (2022) got a better churn prediction. Compared to these models, our study achieved higher accuracy, precision, recall and ROC-AUC and the predictions are incorporated into a DSS. Previous research did not address model integration in-depth or addressed specific key activities in CRM, such as segmentation or feedback only. In our study, we have presented a more elaborate CRM model plan but data privacy issue, real-time responsiveness comprise the future ethical directions. Moreover, the proposed model is compared with existing works as summarized in Table 9.