Comparative Study Of Machine Learning Models (Decision Tree, Random Forest, ANN) For Churn Prediction in Telecom Industry

In the dynamic landscape of the telecom industry, where customer loyalty is constantly challenged by evolving preferences and fierce competition, the ability to predict and prevent customer churn is paramount. This overview sets the stage for a comprehensive exploration of churn prediction models tailored specifically for the telecom sector. Beginning with an analysis of the multifaceted drivers of churn, including service quality, pricing dynamics, and customer satisfaction, this overview highlights the critical importance of proactive churn management strategies for telecom operators. Recognizing the significance of data- driven approaches in addressing this challenge, the overview introduces the concept of churn prediction models as a strategic tool for enhancing customer retention.

The report delves into the development and evaluation of a novel churn prediction model, leveraging advanced machine learning techniques and rich telecom datasets.

In the highly competitive telecom industry, customer retention is critical for sustained profitability. Customer churn, defined as the tendency of subscribers to stop using a telecom service, presents a significant challenge for companies. Predicting churn allows businesses to identify at- risk customers and implement targeted retention strategies. This problem involves developing an effective churn prediction model using Decision Tree, Random Forest, and Artificial Neural Networks (ANN). The objective is to compare these algorithms based on their accuracy, precision, recall, and F1-score to identify the most suitable model for churn prediction.

Detail, in fact, the reasons for customer churn in the telecom sector. Broadcaster service quality, pricing strategy, and customer satisfaction. Build a solid churn predicting mechanism for telecom operators using machine learning algorithms. Help orchestrate the amalgamation of widely varied customer datasets, customer demographics, usage behavior, and service interaction for improved predicted performance of the churn model. Propose strong testing of the churn prediction model using real-time telecom data sets. Show how effective and scalable this devised model is, backing this with comparisons against traditional churn prediction methods and demonstrating its special efficacy. This allows action insights to be provided to the companies for defining the risk of churn and allowing targeted retention policies using interpretability techniques. In turn, all these lead to the firms being able to derive actionable insights for resource optimization, ensuring customer satisfaction before containment, and finally, sustainable growth providing a strong impetus to telecoms to remain competitive.

Churn prediction allows preventive and, most effective intervention in cases of customer attrition that directly affect profitability. More complex machine-learning techniques such as decision tree classifiers, random forests and ANN.

RELATED WORKS

1. Customer churn prediction in telecom sector using Machine Learning techniques: Sharmila K. Wagh et al. (2024) [1]:

This paper studies the churn prediction in the telecom sector comprehensively, using several machine learning algorithms to achieve marketing analysis of churn behaviors. The causes of customer churn should be observed for developing proper retention strategies. The authors reach an accuracy of about 99% with respect to churn predictions based on the decision tree-based classifiers and random forests. To further add to the churn prevention researches, this also includes clinical survival analysis utilizing the Cox proportional hazard model. Data preprocessing and feature selection are two significant contributors to the complete architecture of the proposed system. Regarding customer retention issues, the study issues a warning to service providers that churn provocation is unavoidable; that should be filled with a plea to get a better hold on this aspect. Another future path may also lead towards consideration of some more advanced algorithms which would additionally deal with increasing chances of improving the classification accuracy using deep learning techniques.

2. Customer churn prediction using machine learning approaches: R. Srinivasan et. al (2023) [2]:

In their study "Customer Churn Prediction Using Machine Learning Approaches," the authors examine distinct attempts for dealing with the customer churn problem under the purview of telecom operators' machine learning techniques. It correctly highlights customer identification who is likely going to churn, which may have a great revenue impact. Different approaches involved in algorithm analysis in decision trees and random forest refer to the handling of data imbalance during their churn prediction. Pretty much impressive F1-score based performance evaluation measures were then compared for assessing the performance of the models. Data pre-processing and feature extraction revealed themselves as the most critical areas of action required for improving classification accuracy in churn prediction analyses. Sampling should effectively alleviate the imbalances of datasets and thus favor estimates.

3. Customer churn prediction in the telecom industry using tabular machine learning models: S.S. Poudel et al. (2024) [3]: The paper deals with customer churn prediction in the telecommunication industry, focusing on the early detection of risks of churn to retain the clients. It criticizes the earlier generalized classification methods for not being interpretable enough, which is a necessary requirement for proper decision-making. The research develops explainable machine learning models, and one of the models applied was the Gradient Boosting Machine (GBM), which resulted in 81% accuracy. Visualization techniques, in the form of SHAP plots, are used for the better understanding of the factors of churn. Validating the superior performance is a Wilcoxon signed-rank test. The overall study highlights the importance of local and global explanations that will be used to understand dynamics in churn.

4. Home Appliance Rental Customer Churn Prediction: Youngjung Suh, 2023 [4]: The paper is titled "Customer Churn Prediction in the Home Appliance Rental Business" by Youngjung Suh, published on March 21, 2023, in the Journal of Big Data. It addresses the serious problem of customer churn that poses significant challenges for companies operating in the rental business, specifically those offering home appliances like water purifiers. The research focuses on customer retention strategies because the cost of retaining a current customer is usually cheaper than acquiring a new one. Suh applies machine learning techniques to develop predictive models that evaluate the chances of customer churn based on historical subscription data. The research underlines feature engineering that is informed by domain knowledge in identifying critical predictors of the influence of customer retention. Understanding the churning factors and also the behavior of customers, this paper aims to offer business-friendly actionable insights that would lead toward strategic targeted marketing and care. Hence, finally, it also offers a framework for further advancing knowledge to retain the efforts in the customer dynamics in the rental industry.

5. Customer Churn Prediction Using Machine Learning: Nadipelli Shreshta et al. (2022) [5]:

The paper discusses the development of a churn prediction model that would be able to identify customers likely to churn. The retention of a current customer is better than acquiring a new one since churns affect revenue. This study assessed many machine learning techniques aimed at obtaining highly accurate recognition of churn data like Random Forest, SVM, Logistic Regression, and XGBoost. The proposed model employs ensemble learning to yield better performances in a prediction.

6. A Comparative Analysis of Data Preparation Algorithms for Customer Churn Prediction: A Case Study in the Telecommunication Industry: Kristof Coussement (2016)

[6]: Kristof Coussement, Stefan Lessmann, and Geert Verstraeten analyze the influence of different techniques of data preparation on the performance of the logistic regression churn prediction model implemented with real- world data from one leading European telecommunications provider. The study has global implications since the performance of the optimized logit model is compared with eight advanced data mining techniques. This analysis shows that different choices of data preparation could considerably improve prediction performance, in this case, about 14.5 percent for AUC and 34 percent for top decile lift. Thus, data preprocessing is extremely important in the entire cycle of predictive analytics about customer retention. It closes with managerial implications and recommendations for future research in other business contexts.

7. Machine-learning-based strategy for preventing customer churn in E-commerce: Shobana J et al. (2023). [7]:

This paper presents a machine-learning-based strategy for predicting the customer churn in e-commerce, highlighting that retaining these customers is really important because of stiff competition. The paper talks about the cost incurred when acquiring new customers and features information measuring churn consideration. Using the predictive analysis technique, the SVM implementation will be built into a hybrid recommendation strategy in supplementing retention effort. Management will be able to position the retention offers according to the customer behavior and transaction record. The coverage ratio and precision rate significantly vary with this integrated predictive model. This paper emphasized the understanding of customer attrition as a basis for implementing successful retention strategies. Overall, this paper has proposed a framework of business intelligence aimed at maximizing customer retention in e- commerce.

PROPOSED METHODOLOGY

The Telco Customer Churn dataset used in this research paper is a structured collection of information pertaining to customers of a fictional telecommunications company. It consists of approximately 7,043 records and 21 attributes that capture a comprehensive view of each customer’s demographic profile, service usage patterns, billing information, and churn status. Each entry is uniquely identified by a customerID, and the primary target variable is Churn, which indicates whether a customer has discontinued the service (Yes) or remained (No). The demographic attributes include details such as gender, senior citizen status, partnership status, and whether the customer has dependents. Account-related features include the tenure (number of months with the company), contract type (month-to-month, one-year, or two-year), billing preferences (paperless or not), and payment method. Service-related fields describe the subscription status to various offerings such as phone service, internet service (DSL, fiber optic, or none), and additional internet-based features like online security, backup, device protection, technical support, and streaming services. Furthermore, the dataset includes numerical features such as MonthlyCharges and TotalCharges, which provide insights into customer spending. This diverse range of features makes the dataset highly suitable for exploring patterns associated with customer retention and churn prediction through machine learning techniques.

The proposed methodologies for customer churn forecasting in the telecommunication industry are multi-stage and involve data preprocessing, feature extraction, model training, and evaluation using three machine learning algorithms: Decision Tree, Random Forest, and Artificial Neural Network (ANN). The dataset is preprocessed first to handle missing values, inconsistencies, and converting categorical variables into numerical values using techniques such as label encoding and one-hot encoding. This is followed by scaling features to normalize numeric features so that model efficiency and accuracy are maintained. Following preprocessing, the data is divided into training and testing sets to make measurements of the performance of the model objective.

The second one, Random Forest, is an ensemble of multiple decision trees such that the output prediction is given by a majority vote of individual trees. The method can prevent overfitting and improve generalization performance. The third one uses an Artificial Neural Network (ANN) that mimics the structure of the human brain with neuron layers that are interconnected. The ANN model consists of an input layer, hidden layers, and an output layer, using backpropagation and optimization methods for minimizing prediction error.

All the models—Decision Tree (DT), Random Forest (RF), and Artificial Neural Network (ANN)—are trained on the same dataset and evaluated using standard performance metrics such as accuracy, precision, recall, F1-score, and area under the ROC curve (AUC). In this context, the analysis revealed that tenure and dependents were consistently strong predictors of churn across all models, while gender had minimal impact. Customers with lower tenure (0–12 months) and no dependents were more likely to churn, indicating potential dissatisfaction or lack of service commitment. While DT provided interpretability, RF demonstrated robustness through ensemble learning, and ANN outperformed in capturing complex, non-linear relationships. By these comparisons, the study aims to identify the most suitable algorithm for customer churn prediction and to provide telecom companies with actionable insights for targeted retention strategies.

One The analysis of the tenure characteristic indicated that customer retention is strongly associated with how many months the customer has been a customer of the telecommunication company. Lower-tenure customers, particularly those with 0 to 12 months tenure, had a greater rate of churn than the higher tenures. This indicates that new customers tend to churn off, possibly due to unmet expectations or dissatisfaction. By gender, the distribution of male and female customers was nearly even, and churn rates were not significantly different by gender, suggesting that gender is not strongly predictive of churn. The Dependents feature did show more significant trends — customers with no dependents were more likely to churn than customers with dependents. This may imply that customers with dependents value stability of service more and are less likely to switch providers. Overall, tenure and Dependents were more influential drivers of churn behavior than gender.

In the comparative study of customer churn prediction models— Decision Tree (DT), Random Forest (RF), and Artificial Neural Network (ANN)—the analysis of customer features revealed that tenure and dependents were more influential in predicting churn than gender. All three models identified that customers with lower tenure (0–12 months) had a significantly higher churn rate, indicating that new users are more likely to leave, possibly due to unmet expectations. While the distribution of churn was nearly equal across male and female customers, making gender a weak predictor, the Dependents feature showed clearer patterns: customers without dependents were more likely to churn, suggesting that those with dependents may value service stability more. Among the models, ANN captured subtle non-linear relationships better, RF showed strong performance through ensemble learning, and DT provided clearer interpretability. Despite architectural differences, all three models aligned in highlighting tenure and dependents as key churn indicators.

SYSTEM ARCHITECTURE

Churn shows different stages along the architecture designed to streamline processes from data capture to predicting output. It is based on the IBM Watson dataset, which offers a range of customer attributes: demographics for billing, service usage, and some billing data. In the early stages, data are cleaned for a better output on the model. Major preprocessing steps follow the removal of noise (removal of data points that are non- informative), the normalization of data scales numerical features into a feasible range, and selection of features capable of predicting customer churn. This makes the dataset compatible with machine learning algorithms and, in turn, increases model performance.

From the Initialization phase, the Model Development phase comes on with the Decision Tree, Random Forest, and ANN deployed for this. In this phase, all introduced models train on the data, and by the allowable divisions, they attempt to balance some of the precision recall scores, tuned accuracy to prevent overfit and underfit. Decision Tree offers the simplest yet interpretable rule-based model, Random Forest in turn improves on predictive accuracy by training on multiple trees, while AN is indeed very robust, working in tandem to develop strong relationships over complex high-dimensional datasets.

In the last act of predictive analytics, Fit Predicts the possibilities of the new churn rates drawn from the existing customer data through the predictive models. This affords telecom companies the chance to kickstart and bolster initiatives against churn, instilling loyalty among customers and leaving behind lower rates of customer attrition. A snapshot of this architecture, in a good data-centric regime, fits into the field of churn management in the telecom industry.

Fig 1. System Architecture

Churn shows different stages along the architecture designed to streamline processes from data capture to predicting output. It is based on the IBM Watson dataset, which offers a range of customer attributes: demographics for billing, service usage, and some billing data. In the early stages, data are cleaned for a better output on the model. Major preprocessing steps follow the removal of noise (removal of data points that are non- informative), the normalization of data scales numerical features into a feasible range, From and selection of features capable of predicting customer churn. This makes the dataset compatible with machine learning algorithms and, in turn, increases model performance. The Initialization phase, the Model Development phase comes on with the Decision Tree, Random Forest, and ANN deployed for this. In this phase, all introduced models train on the data, and by the allowable divisions, they attempt to balance some of the precision recall scores, tuned accuracy to prevent overfit and underfit. Decision Tree offers the simplest yet interpretable rule- based model, Random Forest in turn improves on predictive accuracy by training on multiple trees, while AN is indeed very robust, working in tandem to develop strong relationships over complex high-dimensional datasets.

In the last act of predictive analytics, Fit Predicts the possibilities of the new churn rates drawn from the existing customer data through the predictive models. This affords telecom companies the chance to kickstart and bolster initiatives

against churn, instilling loyalty among customers and leaving behind lower rates of customer attrition. A snapshot of this architecture, in a good data-centric regime, fits into the field of churn management in the telecom industry.

The analysis of the tenure characteristic indicated that customer retention is strongly associated with how many months the customer has been a customer of the telecommunication company. Lower-tenure customers, particularly those with 0 to 12 months tenure, had a greater rate of churn than the higher tenures. This indicates that new customers tend to churn off, possibly due to unmet expectations or dissatisfaction. By gender, the distribution of male and female customers was nearly even, and churn rates were not significantly different by gender, suggesting that gender is not strongly predictive of churn. The Dependents feature did show more significant trends — customers with no dependents were more likely to churn than customers with dependents. This may imply that customers with dependents value stability of service more and are less likely to switch providers. Overall, tenure and Dependents were more influential drivers of churn behavior than gender.

In order to pre-process the dataset for machine learning modeling, certain steps of preprocessing were performed on the chosen features: tenure, gender, and Dependents. The tenure feature, being a numeric attribute that represents the number of months the customer has been with the company, was normalized using Min-Max normalization in order to reduce its values to a 0 to 1 range. This assists in enhancing the convergence of learning algorithms, particularly for feature scale-sensitive models such as neural networks. The gender attribute, being a binary categorical feature with values like "Male" and "Female", was label encoded into numeric format, where "Male" was assigned 1 and "Female" 0. Analogously, the Dependents attribute, that indicates if a customer has dependents ("Yes" or "No"), was also label encoded with "Yes" as 1 and "No" as 0. Missing values, if any, were managed suitably—either by deleting rows or imputation by mode (in case of categorical) and median (in case of numerical) values. These preprocessing operations guaranteed that the data was clean, uniform, and ready to be fed into machine learning algorithms.

RESULTS AND ANALYSIS