Telco Customer Churn Prediction

Overview This project builds and evaluates machine learning models to predict customer churn in a telecom dataset. The objective is to identify customers likely to leave the service so that proactive retention strategies can be implemented.

The workflow includes: Data cleaning and preprocessing Feature encoding and scaling Model training

Evaluation and comparison Two models are implemented: Logistic Regression (baseline) Random Forest Classifier (improved model)

Dataset The dataset contains telecom customer records including demographics, billing information, and service usage. The target variable is: Churn → Whether a customer leaves the service.

Train / Validation Split Method The dataset is divided into training and validation sets using an 80/20 stratified split. 80% → Training data 20% → Validation data

Stratified sampling ensures that the proportion of churn vs non-churn customers remains consistent across both sets.

Implementation concept: Random seed fixed for reproducibility Stratification applied to preserve class balance

This split allows reliable performance evaluation while maximizing training data.

Model Training:

Logistic Regression (Baseline):

Numerical features scaled Provides a linear reference model

Random Forest (Improved Model):

Ensemble tree-based model Handles nonlinear feature interactions Tuned hyperparameters for improved generalization

Evaluation Metrics: Model performance is evaluated using multiple classification metrics to capture different aspects of prediction quality: Accuracy — overall correctness

Precision — correctness of churn predictions

Recall — ability to detect churn cases

F1 Score — balance between precision and recall

ROC-AUC — ranking quality of predictions

Confusion Matrix — class-wise prediction breakdown

Using multiple metrics ensures balanced evaluation beyond simple accuracy.

Results: Logistic Regression (Baseline) Accuracy: ~0.80 F1 Score: ~0.61 ROC-AUC: ~0.84

Random Forest (Improved Model) Accuracy: ~0.83 F1 Score: ~0.67 ROC-AUC: ~0.86

Best Result: The Random Forest model achieved the best overall performance.

Key improvements: Higher prediction accuracy Better churn detection (recall) Stronger F1 score balance Improved ROC-AUC This indicates that the ensemble model captures complex patterns in customer behavior more effectively than the baseline linear model.

Conclusion: This project demonstrates a complete machine learning pipeline: Structured preprocessing Model comparison Robust evaluation

The Random Forest model provides the strongest predictive performance and serves as the recommended approach for churn prediction in this dataset.

Additional Analysis & Visualizations

Exploratory Data Analysis (EDA) Visualizations

Churn Distribution Analysis

The dataset shows a ~26.5% churn rate, with the majority of customers (73.5%) retaining the service. This class imbalance is important to consider during model training.

Statistical Correlations

Correlation analysis reveals strong relationships between key features:

• Tenure and TotalCharges: 0.83 correlation (customers stay longer and pay more)
• MonthlyCharges and TotalCharges: 0.65 correlation
• These numerical features are critical predictors of customer behavior

Model Performance Visualizations

Confusion Matrices Analysis

The confusion matrices show the trade-off between model types:

Logistic Regression (Baseline):

• True Negatives: 925 (correctly predicted non-churners)
• False Positives: 110 (incorrectly predicted churners)
• False Negatives: 162 (missed churners - critical misses)
• True Positives: 212 (correctly identified churners)

Random Forest (Improved):

• True Negatives: 941 (improved non-churner detection)
• False Positives: 94 (fewer false alarms)
• False Negatives: 179 (slightly more missed churners)
• True Positives: 195 (fewer caught churners, but balanced)

ROC Curve Comparison

Both models achieve similar ROC-AUC scores:

• Logistic Regression: AUC = 0.8416 (blue curve)
• Random Forest: AUC = 0.8439 (green curve)
• Random Classifier Baseline: AUC = 0.5 (dotted line)

The curves show excellent discrimination ability for both models, with only marginal difference between them.

Feature Importance Analysis

The Random Forest model reveals the most important features for churn prediction (Top 15):

1. Tenure (0.2095) - Strongest predictor; longer tenure = lower churn
2. TotalCharges (0.1339) - Cumulative spending indicates loyalty
3. MonthlyCharges (0.1045) - High charges increase churn risk
4. InternetService_Fiber optic (0.0797) - Fiber optic customers have different churn patterns
5. PaymentMethod_Electronic check (0.0683) - Payment method affects retention
6. Contract_Two year (0.0657) - Long-term contracts reduce churn
7. OnlineSecurity_Yes (0.0399) - Services reduce churn risk
8. Contract_One year (0.0374) - Medium-term contracts provide some protection
9. TechSupport_Yes (0.0240) - Support services improve retention
10. PaperlessBilling_Yes (0.0196) - Administrative preferences matter

Key Insights from Analysis

1. Contract Type Impact: Customers with month-to-month contracts show significantly higher churn than those with annual or multi-year contracts
2. Tenure Effect: The first 12 months are critical; early intervention is crucial
3. Service Adoption: Online security and tech support services strongly correlate with lower churn
4. Pricing Strategy: Very high monthly charges correlate with increased churn risk
5. Customer Segmentation: Internet service type (Fiber optic vs DSL) shows distinct churn patterns

Recommendations for Implementation

1. Proactive Retention for New Customers: Focus retention efforts on customers in their first 12 months
2. Contract Optimization: Incentivize customers to upgrade from month-to-month contracts
3. Service Bundling: Promote online security and tech support services as churn reduction tools
4. Pricing Adjustment: Review pricing for high-paying customers to improve value perception
5. Targeted Offers: Use model predictions to identify at-risk customers and deploy personalized retention campaigns
6. Internet Service Focus: Develop fiber optic specific retention strategies

Model Performance Metrics Summary

Both models achieved comparable performance on validation data:

• Accuracy: ~80.7% - Correctly classifies 4 out of 5 customers
• Precision: 66-67% - When model predicts churn, it's right 2/3 of the time
• Recall: 52-57% - Can identify 50-57% of actual churners
• F1-Score: 0.59-0.61 - Balanced performance metric

The trade-off between models: Random Forest shows slightly better overall ROC-AUC but similar accuracy, suggesting both approaches are viable depending on business priorities (minimize false positives vs. catch more churners).