Explainable AI-Based Deep Learning System for Predicting Customer Churn in Telecommunication Industry

Abstract

Customer churn is a major challenge for the telecommunication industry as the loss of customers impacts revenue and business growth. Early identification of customers who are likely to churn allows telecom providers to develop effective customer retention strategies and increase customer satisfaction. Traditional machine learning approaches often fail to capture complex customer behavior patterns and offer limited interpretability in their predictions. To overcome these challenges, this project proposes an Explainable AI-Based Deep Learning System for prediction of customer churn in telecommunication industry. The system uses IBM Telco customer churn dataset and uses exhaustive data preprocessing techniques such as data cleaning, label encoding, feature scaling and class balancing using SMOTEENN. It employs a hybrid deep learning architecture called ChurnNet comprising a 1D Convolutional Neural Network (1D-CNN), Residual Blocks, Channel Attention, and Spatial Attention mechanisms to learn complex customer behavioural patterns and accurately predict churn probability. In order to enhance credibility and clarity, the suggested system incorporates Explainable Artificial Intelligence (XAI) methods like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations) to provide transparency. SHAP performs global feature importance analysis, and LIME provides local explanations for each prediction, allowing users to understand the crucial features influencing the churn decision. The experimental results show that the proposed system has high prediction accuracy, reliability and interpretability, and thus it is a valuable decision-support tool for telecom organizations that aim to reduce customer attrition and improve retention strategies.

Introduction

The text presents a customer churn prediction system for the telecommunications industry based on a hybrid deep learning architecture called ChurnNet. Customer churn is a major business challenge because retaining existing customers is generally more cost-effective than acquiring new ones. Although telecom companies collect large amounts of customer data, identifying customers likely to leave is difficult because of complex and diverse usage behaviours.

The proposed system combines deep learning with Explainable Artificial Intelligence (XAI) to provide both accurate predictions and understandable explanations. ChurnNet uses a 1D-CNN, Residual Blocks, Channel Attention, and Spatial Attention to learn complex customer behaviour patterns. SHAP and LIME are then used to explain why the model predicts a particular customer as likely or unlikely to churn.

Literature Survey

Previous research has applied CNNs, BiLSTM, ConvLSTM, attention mechanisms, ensemble models, and optimization techniques to customer churn prediction. These approaches have improved prediction accuracy but often increase model complexity, computational requirements, and lack of interpretability.

Other studies have addressed important issues such as:

• Class imbalance between churn and non-churn customers.
• Feature selection and data transformation.
• Identification of important customer characteristics.
• Just-in-time churn prediction for proactive customer retention.

However, many advanced deep-learning models remain difficult to interpret. The proposed ChurnNet addresses this limitation by combining deep feature learning, class balancing, attention mechanisms, and XAI in one framework.

Proposed Methodology

The system follows several major stages:

1. Data Collection:
   The IBM Telco Customer Churn Dataset is used, containing 7,043 customer records and 21 variables, including demographic information, services, billing information, tenure, charges, and churn status.

2. Data Preprocessing:
   Missing or invalid values are handled, irrelevant attributes such as customer ID are removed, categorical variables are converted into numerical form, and numerical features are scaled.

3. Class Balancing:
   SMOTEENN is applied to address the imbalance between churn and non-churn customers. SMOTE generates synthetic minority-class samples, while ENN removes noisy or inappropriate samples.

4. ChurnNet Model:
   The processed data is passed to a hybrid deep-learning model containing:

   • 1D-CNN for extracting important feature patterns.
   • Residual Blocks for deeper feature learning.
   • Channel Attention for emphasizing important feature channels.
   • Spatial Attention for focusing on significant feature regions.

5. Training and Prediction:
   The model is trained using stratified cross-validation. For new customers, the system predicts whether they are likely to churn and produces a churn probability/risk score.

6. Explainable AI:
   SHAP provides an overall understanding of which features influence churn predictions, while LIME explains the reasons behind individual customer predictions.

7. Performance Evaluation:
   The model is evaluated using Accuracy, Precision, Recall, F1-Score, AUC-ROC, and Matthews Correlation Coefficient (MCC).

Experimental Results

The proposed ChurnNet model achieved strong performance:

These results indicate that ChurnNet can accurately distinguish between customers who are likely to churn and those who are likely to remain. The high recall is particularly important because it means the system can identify a large proportion of actual churn customers, allowing telecom companies to take preventive retention measures.

Conclusion

The proposed system for Customer Churn Prediction, based on Explainable AI, effectively showcases the fusion of deep learning and explainable artificial intelligence for predicting churn in the telecommunications sector. The ChurnNet model, incorporating 1D-CNN, Residual Blocks, and Attention Mechanisms, adeptly captures intricate customer behavior patterns to deliver precise churn forecasts. Mitigating class imbalance through SMOTEENN and ensuring robust model assessment via 10-Fold Cross Validation are integral components of this system. The incorporation of SHAP and LIME bolsters model transparency by elucidating the variables influencing churn predictions, thereby fostering comprehension and trust in the system. Moreover, the user-friendly Streamlit-based interface facilitates interactive customer analysis and prediction visualization. Empirical findings underscore the system\'s robust performance, evidenced by elevated Accuracy, Precision, Recall, F1- Score, MCC, and AUC-ROC metrics. In summary, this system offers a precise, dependable, and interpretable solution for customer churn prediction, empowering telecom providers to pinpoint high-risk customers and deploy effective retention strategies. Prospective refinements could encompass leveraging larger datasets, advanced deep learning frameworks, and realtime deployment to enhance predictive efficacy.

References

[1] S. Saha, C. Saha, M. M. Haque, M. G. R. Alam, and A. Talukder, “ChurnNet: Deep Learning Enhanced Customer Churn Prediction in Telecommunication Industry,” IEEE Access, vol. 12, pp. 4471–4484, 2024. DOI: 10.1109/ACCESS.2024.3349950. [2] C. Yang, G. Xia, L. Zheng, X. Zhang, and C. Yu, “Customer Churn Prediction Based on Coordinate Attention Mechanism with CNN-BiLSTM,” Electronics, vol. 14, no. 10, p. 1916, 2025. DOI: 10.3390/electronics14101916. [3] H. Zhang and W. Zhang, “Application of GWO-attention-ConvLSTM model in customer churn prediction and satisfaction analysis in customer relationship management,” Heliyon, vol. 10, no. 17, p. e37229, 2024. DOI: 10.1016/j.heliyon.2024.e37229. [4] Y. Liu, M. Shengdong, G. Jijian, and N. Nedjah, “Intelligent prediction of customer churn with a fused attentional deep learning model,” Mathematics, vol. 10, no. 24, p. 4733, 2022. [5] N. Jajam, N. P. Challa, K. S. L. Prasanna, and C. H. V. S. Deepthi, “Arithmetic optimization with ensemble deep learning SBLSTM-RNN-IGSA model for customer churn prediction,” IEEE Access, vol. 11, pp. 93111–93128, 2023. [6] S. Loukili et al., “Enhancing Customer Retention through Deep Learning and Imbalanced Data Techniques,” Iraqi Journal of Science, vol. 65, no. 5, pp. 2853–2866, 2024. [7] A. Khattak, Z. Mehak, H. Ahmad, M. U. Asghar, M. Z. Asghar, and A. Khan, “Customer churn prediction using composite deep learning technique,” Scientific Reports, vol. 13, no. 1, Art. no. 17294, 2023. [8] A. Amin, S. Anwar, A. Adnan, M. Nawaz, N. Howard, J. Qadir, A. Hawalah, and A. Hussain, “Comparing oversampling techniques to handle the class imbalance problem: A customer churn prediction case study,” IEEE Access, vol. 4, pp. 7940–7957, 2016. [9] J. K. Sana, M. Z. Abedin, M. S. Rahman, and M. S. Rahman, “A novel customer churn prediction model for the telecommunication industry using data transformation methods and feature selection,” PLOS ONE, vol. 17, no. 12, Art. no. e0278095, 2022. [10] L. AlShourbaji, N. Helian, Y. Sun, A. G. Hussien, L. Abualigah, and B. Elnaim, “An efficient churn prediction model using gradient boosting machine and metaheuristic optimization,” Scientific Reports, vol. 13, no. 1, p. 14441, 2023. [11] A. Amin, F. Al-Obeidat, B. Shah, M. A. Tae, C. Khan, H. U. R. Durrani, and S. Anwar, “Just-in-time customer churn prediction in the telecommunication sector,” The Journal of Supercomputing, vol. 76, no. 6, pp. 3924–3948, 2020. [12] I. Ullah, B. Raza, A. K. Malik, M. Imran, S. U. Islam, and S. W. Kim, “A churn prediction model using random forest: Analysis of machine learning techniques for churn prediction and factor identification in telecom sector,” IEEE Access, vol. 7, pp. 60134–60149, 2019.

Copyright

Copyright © 2026 Kasam Teja, M. Arathi. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.