Customer Churn Prediction

An end-to-end machine learning project to predict customer churn using the Telco dataset from Kaggle.

Dataset

• Source: Telco Customer Churn (11.1.3+)
• Contains information about a fictional telco company that provided home phone and internet services to 7043 customers in California in Q3. It indicates which customers have left, stayed, or signed up for their service.

Exploratory Data Analysis (EDA)

Performed exploratory data analysis churn_EDA.ipynb to understand customer behavior patterns leading to churn. Key insights include:

1. Class Imbalance:
   • 73.5% customers did not churn (No)
   • 26.5% customers churned (Yes)
2. Demographics & Churn:
   • Senior citizens have a higher likelihood of churn
   • Customers with dependents are less likely to churn
   • Gender has no major impact on churn
3. Behavioral Patterns:
   • Month-to-month contract users churn more often than those with 1 or 2-year contracts
   • Customers using Bank Withdrawal as payment method churn more than others
   • Users with low tenure and low total charges churn more frequently
   • Fiber optic internet users have a higher churn rate compared to DSL or Cable

Data Cleaning & Preprocessing

• Dropped irrelevant or redundant features like customer_id, under_30, quarter, churn_category, and churn_reason in churn_DataPreProcessing.ipynb
• Converted all binary features (Yes/No) and gender (Male/Female) to 1/0
• One-hot encoded multi-class categorical features:
  • internet_type
  • contract
  • payment_method
• Created two separate datasets:
  • linear_data.csv: for linear models (one-hot encoded with `drop_first=True)
  • tree_data.csv: for tree-based models (full one-hot encoding)

Handling Class Imbalance

• Observed class imbalance: ~73% No Churn vs. ~27% Yes Churn in churn_resampling.ipynb
• Compared and implemented 3 resampling strategies (on training data only):
  • Random UnderSampling: reduced majority class
  • Random OverSampling: duplicated minority class
  • SMOTE: used to generate synthetic samples for the minority class
• Selected SMOTE as the preferred method for modeling

Model 1: Logistic Regression (Baseline)

This is the first of several models used to predict customer churn.

• Trained on a SMOTE-balanced dataset to address class imbalance.
• Dropped multicollinear features based on a correlation matrix
• Used Logistic Regression as a baseline linear model due to its interpretability and strong performance

Evaluation on Real-World Test Set

• ROC AUC Score: 0.97 (excellent class separability)
• Accuracy: 93%
• Precision / Recall (No Churn): 95% / 95%
• Precision / Recall (Churn): 87% / 87%
• F1-score: Balanced across both classes

Model 2: Ridge Classifier (L2 Regression)

• Built on logistic regression by introducing L2 regularization
• Helped reduce potential multicollinearity and slightly improved performance
• Precision for churn class improved from 0.87 -> 0.92
• Overall accuracy increased from 93% -> 95%
• Retained all features but shrunk coefficient magnitudes for better generalization

Model 3: Lasso Regression (L1 Regression)

• Applied L1 regularization to perform automatic feature selection
• Shrunk less important feature coefficients to zero
• Highlighted the most impactful features for churn:
  • satisfaction_score
  • online_security
  • phone_service
  • internet_service
• Improved model interpretability and reduced complexity
• Ideal when the goal is to identify key drivers of churn

Conclusion: Linear Models

1. All three models (Logistic, Ridge, Lasso) performed well on the SMOTE-balanced dataset
2. Logistic Regression: Strong baseline with excellent ROC AUC (0.97)
3. Ridge Classifier: Slightly better precision and acuuracy, useful when all features matter
4. Lasso Regression: Best for understanding which features matter most, by reducing noise
5. These models laid the foundation for building interpretable churn prediction systems

Model 4: Decision Tree Classifier

• Implemented a Decision Tree Classifier to predict customer churn
• Utilized the SMOTE-balanced dataset to address class imbalance
• Achieved the following performance metrics on the test set:
  • Accuracy: 95%
  • Precision: 0.96 (No Churn), 0.90 (Churn)
  • Recall: 0.96 (No Churn), 0.90 (Churn)
  • F1-Score: 0.96 (No Churn), 0.90 (Churn)
  • ROC AUC Score: 0.93
• The model demonstrated balanced performance across both classes.

Model 5: RandomForest Classifier

• Trained a Random Forest model to predict customer churn
• Achieved the strongest performance among all the models so far:
  • Accuracy: 95%
  • Precision: 0.98 (Churn), 0.95 (No Churn)
  • Recall: 0.84 (Churn), 0.99 (No Churn)
  • F1-score: 0.91 (Churn), 0.97 (No Churn)
  • ROC AUC Score: 0.98
• Very few false positives and high class separation, making it ideal for customer churn prioritization.

Model 6: XGBoost Classifier

• Trained an XGBoost Classifier using the cleaned, imbalanced tree dataset.
• Achieved the best model performance across all models evaluated:
  • Accuracy: 96%
  • Precision: 0.95 (Churn), 0.96 (No Churn)
  • Recall: 0.90 (Churn), 0.98 (No Churn)
  • F1-score: 0.92 (Churn), 0.97 (No Churn)
  • ROC AUC Score: 0.99
• Excellent balance between precision and recall, making it highly suitable for churn prediction tasks.

Conclusion: Tree-based Models

1. Tree-based models (Decision Tree, Random Forest, XGBoost) performed well without requiring SMOTE balancing
2. Decision Tree Classifier: Provided a good baseline but slightly overfitted on training data
3. Random Forest Classifier: Improved stability and performance, especially in precision (0.98) with ROC AUC of 0.98
4. XGBoost Classifier: Best overall performance with highest accuracy (0.96) and ROC AUC (0.99), showing excellent generalization
5. These models are powerful for capturing non-linear relationships and identifying key churn patterns in the data

Final Thoughts

This project successfully built an end-to-end machine learning pipeline to predict customer churn using a realistic Telco dataset. Through thorough exploratory data analysis, strategic preprocessing, handling class imbalance, and evaluating multiple linear and tree-based models, I was able to:

• Achieve a high-performing XGBoost model with 96% accuracy and 0.99 ROC AUC.
• Understand the key drivers of customer churn such as contract type, tenure, internet service, and payment method.
• Compare models not only on performance but also on interpretability and feature influence.

This pipeline can now be extended or deployed for:

• Real-time churn prediction on new customer data
• Business dashboards highlighting high-risk customers
• A/B testing retention strategies based on churn scores