Customer Churn Prediction

Project Overview

Customer churn is one of the most critical challenges faced by subscription-based businesses. Retaining existing customers is significantly more cost-effective than acquiring new ones, making churn prediction a valuable business application of machine learning.

This project develops and evaluates multiple machine learning models to predict customer churn using the Telco Customer Churn dataset. The solution incorporates a complete machine learning pipeline, including data preprocessing, feature engineering, model comparison, hyperparameter tuning, threshold optimization, and model interpretability.

---

Business Objective

The objective of this project is to identify customers who are likely to discontinue a service so that retention strategies can be applied proactively.

Business Benefits

• Reduce customer attrition
• Improve customer lifetime value
• Enable targeted retention campaigns
• Optimize marketing and customer support efforts

---

Dataset

The dataset contains customer demographic information, account details, service subscriptions, and billing information.

Target Variable

Variable	Description
Churn	Whether a customer left the company (Yes/No)

Dataset Characteristics

• Records: 7,032 customers
• Features: 19 predictive features
• Target: Binary Classification
• Churn Rate: ~26.6%

---

Application Screeshots

---

Project Workflow

1. Data Cleaning

• Converted TotalCharges to numeric format
• Removed invalid records containing missing values
• Selected relevant numerical and categorical features

2. Data Preprocessing

Implemented a production-ready preprocessing pipeline using:

• Missing value imputation
• Standard scaling for numerical features
• One-Hot Encoding for categorical variables
• ColumnTransformer for automated feature transformation

3. Train-Test Split

• Stratified train-test split
• 70% Training Data
• 30% Testing Data

4. Model Development

The following models were trained and evaluated:

1. Logistic Regression
2. Random Forest Classifier
3. XGBoost Classifier

5. Hyperparameter Tuning

Random Forest was optimized using:

• GridSearchCV
• Cross-validation
• ROC-AUC as the optimization metric

6. Model Explainability

Implemented SHAP (SHapley Additive Explanations) to understand feature contributions and improve model interpretability.

---

Model Performance

Model	ROC-AUC	F1 Score	Precision	Recall
Logistic Regression	0.86	0.64	0.52	0.84
Random Forest (Tuned)	0.85	0.65	0.56	0.78
XGBoost	0.84	0.62	0.52	0.76

---

Final Model

Tuned Random Forest Classifier

The tuned Random Forest model was selected as the final model due to its strong balance between Precision, Recall, F1 Score, and interpretability.

Best Hyperparameters

{
    "max_depth": 10,
    "min_samples_leaf": 5,
    "min_samples_split": 2,
    "n_estimators": 400
}

Final Performance

Metric	Score
ROC-AUC	0.845
Accuracy	0.772
Precision	0.55
Recall	0.73
F1 Score	0.63

---

Key Drivers of Churn

Feature importance analysis identified the following major churn indicators:

• Customer Tenure
• Contract Type
• Total Charges
• Monthly Charges
• Internet Service Type
• Online Security
• Technical Support
• Payment Method

SHAP analysis was used to validate and interpret these findings.

---

Threshold Optimization

Instead of relying solely on the default classification threshold (0.50), multiple thresholds were evaluated to analyze the trade-off between:

• Precision
• Recall
• F1 Score

This enables business stakeholders to select an operating threshold based on organizational priorities.

---

Technologies Used

Programming Language

• Python

Data Analysis

• Pandas
• NumPy

Data Visualization

• Matplotlib
• Seaborn

Machine Learning

• Scikit-Learn
• XGBoost

Model Explainability

• SHAP

Model Persistence

• Joblib

---

Project Structure

Customer-Churn-Prediction/
│
├── data/
│   └── Telco_Customer_Churn.csv
│
├── notebooks/
│   └── churn_prediction.ipynb
│
├── artifacts/
│   ├── logistic_model.pkl
│   ├── rf_model.pkl
│   └── xgb_model.pkl
│
├── requirements.txt
├── README.md
└── LICENSE

---

Key Achievements

• Built an end-to-end machine learning pipeline for customer churn prediction.
• Implemented automated preprocessing using Scikit-Learn Pipelines.
• Compared multiple classification algorithms.
• Performed hyperparameter optimization using GridSearchCV.
• Applied threshold tuning for business-driven decision making.
• Used SHAP for model interpretability and feature impact analysis.
• Saved trained models for future deployment.

---

Author

Priyesh Kumar Kashyap

Machine Learning Engineer | Data Science Enthusiast