SyriaTel Customer Churn Prediction Project

This project applies machine learning to predict customer churn for SyriaTel, a leading telecommunications provider. It follows the full data science pipeline from data exploration and preprocessing to model development and actionable recommendations to enable the business to proactively identify at-risk customers and reduce churn.

Business Problem

Customer churn presents a critical threat to SyriaTel’s revenue and customer base. Retaining existing customers is far more cost-effective than acquiring new ones. High churn leads to revenue loss, increased customer acquisition costs, and lower lifetime value.

SyriaTel seeks a data-driven churn prediction model that can:

• Identify customers likely to churn
• Uncover behavioral and service-related churn drivers
• Guide personalized, cost-effective retention actions

Project Objectives

This project aims to support SyriaTel’s retention strategy by answering:

1. Which customer behaviors and service patterns predict churn?
2. Can we identify segments at higher risk of leaving?
3. What proactive actions can reduce churn based on model insights?

Dataset

• Source: SyriaTel Telecom Churn Dataset (Kaggle)

Data Cleaning & Preprocessing

• Removed unnecessary columns like phone number, area code and state
• Converted categorical variables that is international plan, voice mail plan to numeric
• Checked and confirmed absence of null values
• Created dummy variables for categorical features

Exploratory Data Analysis (EDA)

EDA was conducted to understand distributions, spot imbalances, and detect patterns:

• Target Variable: Dataset is moderately imbalanced, with ~85.5% non-churned and ~14.5% churned customers.
• Categorical Features: Strong relationship found between churn and features like:
  • International plan: Users with international plans had significantly higher churn rates.
  • Customer service calls: More calls often correlated with dissatisfaction.
• Numerical Features:
  • High day charge and long day minutes were associated with increased churn risk.
  • Visualizations (e.g., histograms and boxplots) revealed useful churn signals.
• Correlation Heatmap: Identified multicollinearity between features such as total_day_minutes and total_day_charge.

Feature Engineering

• Multicollinearity Check:
  • Applied Variance Inflation Factor (VIF) to drop highly correlated features for example dropped total_day_charge in favor of total_day_minutes.
• Class Imbalance Handling:
  • Applied SMOTE (Synthetic Minority Over-sampling Technique) to balance the target variable.
• Encoding:
  • Label encoding was used for binary categorical features that is churn.

Tableau Dashboard:
View interactive visual insights here:
Tableau Dashboard

Modelling

The following models were trained and evaluated:

Model	Accuracy	Recall	Precision	ROC AUC
Logistic Regression	0.681	0.680	0.266	0.768
Decision Tree	0.882	0.732	0.573	0.837
Tuned Decision Tree	0.873	0.680	0.550	0.821
Random Forest	0.882	0.608	0.590	0.860
Tuned Random Forest	0.894	0.660	0.627	0.877

Final Model: Decision Tree with Threshold Optimization

While Random Forest slightly outperformed in AUC, the Decision Tree model was selected for its:

• High recall (important for identifying churners),
• Competitive performance,
• Interpretability.

Threshold Optimization Results:

• Optimal Threshold: 0.421
• Default Recall: 0.732
• Improved Recall: 0.753
• Improved F1 Score: 0.655

Confusion Matrix (Optimized Threshold): [[517 53] [ 24 73]]

Classification Report (Optimized Threshold):

      precision    recall  f1-score   support
   0       0.96      0.91      0.93       570
   1       0.58      0.75      0.65        97

Financial Impact Estimation

• Annual Revenue at Risk: $59,243.04
• False Positive Costs: $2,650.00
• Missed Churn Revenue: $19,056.84
• Recall Rate After Threshold Tuning: 75.3%

Key Visualizations

1. Target Variable Distribution
   Shows the imbalance in churn vs non-churn customers.

2. Top 10 Features Most Correlated with Churn
   Highlights which variables have strongest relationships with churn .

3. Top 10 Feature Importances from the Decision Tree
   Demonstrates which features had the most predictive power in the Decision Tree model.

Recommendations

1. Target Customers Without International Plans
   Offer bundled or discounted international plans to reduce churn in this group.

2. Address High Customer Support Interaction Early
   Use follow-ups and satisfaction surveys to retain users with >3 support calls.

3. Personalize Strategies for Heavy Day/Night Callers
   Loyalty bonuses or discounted rates can reduce churn for high-usage segments.

4. Leverage Churn Predictions in CRM
   Integrate the model into CRM workflows to trigger personalized offers in real time.

5. Prioritize High-Risk Users
   Focus retention resources on customers with high day usage and support calls.

6. Regional & Onboarding Strategy
   Focus on churn-heavy states like NJ, CA, TX with improved onboarding experiences.

7. Incentivize High-Charge Users
   Implement personalized or tiered pricing for customers with high monthly bills.

Limitations

1. SMOTE May Cause Overfitting
   Synthetic samples may reduce generalization without proper validation.

2. Model Diversity
   Only three models were explored; more could boost performance.

3. Support Call Outcome Missing
   Call resolution data could improve prediction accuracy.

4. Model Drift Over Time
   Telecom behavior evolves — periodic retraining will be needed.

Next Steps

1. Try Advanced Models
   Explore XGBoost, LightGBM, and neural networks to enhance accuracy.

2. A/B Test Retention Tactics
   Validate model-informed strategies with real-world experiments.