Telecom Customer Churn Prediction

Overview

This project is a comprehensive Data Science portfolio piece focused on predicting telecom customer churn through advanced machine learning methods. The objective is to develop robust, scalable, and interpretable predictive models to proactively identify at-risk customers, enhancing customer retention strategies.

The analysis builds upon the foundational approach presented by Lalwani et al. (2022), addressing the challenges of imbalanced datasets in churn prediction using advanced resampling techniques (including undersampling and oversampling with SMOTE) and enhanced evaluation metrics. Key methodological highlights include rigorous feature selection using the Gravitational Search Algorithm (GSA), extensive hyperparameter tuning to optimize model performance, and comprehensive comparative analysis across various machine learning models and dataset variations.

📂 Project Structure

The project is organized into two main parts:

Part I: Data Exploration & Preprocessing

1. Introduction: Objectives and scope of churn prediction.
2. Dataset Acquisition & Structure: Data sourcing and preliminary assessment.
3. Exploratory Data Analysis (EDA): Visualization and feature distribution analysis.
4. Data Preprocessing: Handling missing data, transforming categorical features, and data cleaning.

Part II: Modeling & Evaluation

1. Feature Selection & Preprocessing

   • Feature selection using Gravitational Search Algorithm (GSA) and statistical methods.
   • Data balancing with undersampling and SMOTE.

2. Model Training & Hyperparameter Tuning

   • Training supervised learning algorithms (Logistic Regression, Decision Tree, KNN, Random Forest, Naïve Bayes, XGBoost, CatBoost).
   • Hyperparameter optimization for improved predictions.

3. Testing & Model Comparison

   • Evaluation of predictive performance using multiple metrics (Accuracy, Precision, Recall, F1-score, AUC-ROC).
   • Comparing model effectiveness across different feature sets and data balancing methods.

4. Final Recommendation

   • Selection of the best-performing, most interpretable predictive model.
   • Insights into how feature selection and balancing methods affect performance.

⚖ Model and Dataset Comparison

Dataset Variants Evaluated:

• df_original: Complete dataset with all initial features.
• df_gsa: Features selected through GSA optimization.
• df_selected: Features chosen by multiple methods (Correlation, Mutual Information, t-test, Random Forest, XGBoost).
• df_selected_multi: Dataset derived from df_selected after removing highly correlated features.

📈 Key Insights:

Dataset Variant	Number of Features	Performance	Interpretability	Recommended
df_original	31	Moderate	Low	❌
df_gsa	17	Moderate	Moderate	⚠️
df_selected	24	High	Good	✅
df_selected_multi	19	High	Very High	⭐

Recommendation: The dataset df_selected_multi offers the optimal balance of predictive accuracy, computational efficiency, and interpretability, making it ideal for actionable churn prediction strategies.

🤖 Machine Learning Models Implemented

• Logistic Regression
• Decision Tree (CART)
• K-Nearest Neighbors (KNN)
• Random Forest
• Naïve Bayes (Gaussian)
• XGBoost
• CatBoost

📊 Evaluation Metrics

The following performance metrics were used to evaluate the models:

• Accuracy
• Precision
• Recall (Sensitivity)
• F1-Score
• AUC-ROC
• False Negative Rate (Type I Error)
• False Positive Rate (Type II Error)

🛠 Technologies & Tools Used

• Programming Language: Python (Pandas, NumPy, Scikit-learn, XGBoost, CatBoost)
• Visualization: Matplotlib, Seaborn
• Machine Learning Tools: GridSearchCV, Stratified K-Fold Cross-Validation
• Data Preprocessing: SMOTE, StandardScaler, Label Encoding, One-Hot Encoding, Gravitational Search Algorithm (GSA)

📂 Dataset Information

• Source: Telco Customer Churn Dataset by Blastchar (Kaggle)
• Number of Samples: 7,043
• Features: Customer demographics, account details, service subscriptions, billing information.
• Target Variable: Churn (Yes or No)

📅 Conclusion

Accurately predicting customer churn is crucial for telecom providers aiming for proactive retention. This project emphasizes the importance of feature selection and proper handling of imbalanced datasets to achieve highly effective churn prediction. The chosen model provides robust predictive performance alongside meaningful interpretability, enabling efficient, targeted customer retention initiatives.

📚 References

• Lalwani, P., Mishra, M. K., Chadha, J. S., & Sethi, P. (2022). Customer churn prediction system: A machine learning approach. Computing, 104, 1497–1516. doi.org/10.1007/s00607-021-00908-y
• Kaggle - Telco Customer Churn Dataset by Blastchar. Telco Customer Churn Dataset

⚠️ Important Note

This project was developed as part of a Data Science portfolio to explore the application of machine learning techniques for telecom customer churn prediction. The results are not intended to be used as definitive business strategies or a formal scientific study, but rather as a demonstration of data-driven approaches to customer retention.