Customer Churn Analysis

A machine learning project to predict customer churn in the telecom industry using Logistic Regression, Random Forest, and Support Vector Machine (SVM).

Table of Contents

• Overview
• Dataset
• Project Structure
• Tech Stack
• How to Run
• Models Used
• Results
• Visualizations
• Future Improvements
• Team

---

Overview

Customer churn — when a customer stops using a service — is a critical problem for telecom companies. Retaining existing customers is significantly cheaper than acquiring new ones. This project builds a binary classification pipeline that predicts whether a customer will churn based on their usage patterns and account information.

Key goals:

• Clean and preprocess real-world telecom data
• Train and compare three ML models
• Evaluate using multiple metrics (Accuracy, F1, ROC-AUC, Log-Loss)
• Visualize EDA findings and model performance

---

Dataset

Source: Telco Customer Churn — Kaggle

Property	Value
Records	7,043 customers
Features	19 (numerical + categorical)
Target	Churn (1 = churned, 0 = stayed)
Churn Rate	~26.5%

Key features include: tenure, MonthlyCharges, TotalCharges, Contract type, InternetService, PaymentMethod, SeniorCitizen, and more.

---

Project Structure

churn-prediction/
│
├── churn_prediction.py          # Main script — full ML pipeline
├── model_summary.csv            # Model metrics comparison table
│
├── plots/
│   ├── 01_churn_distribution.png
│   ├── 02_boxplots.png
│   ├── 03_correlation_heatmap.png
│   ├── 04_confusion_matrices.png
│   ├── 05_roc_curve_comparison.png
│   ├── 06_learning_curve_accuracy.png
│   ├── 07_learning_curve_logloss.png
│   ├── 08_feature_importance_lr.png
│   ├── 09_feature_importance_rf.png
│   ├── 10_scatter_feature_importance.png
│   └── 11_model_summary_table.png
│
└── README.md

---

Tech Stack

Library	Purpose
pandas	Data loading and manipulation
numpy	Numerical operations
matplotlib	Plotting and visualizations
seaborn	Statistical visualizations
scikit-learn	ML models, preprocessing, evaluation

---

How to Run

1. Clone the repository

git clone https://github.com/sanikatare/churn-prediction.git
cd churn-prediction

2. Install dependencies

pip install pandas numpy matplotlib seaborn scikit-learn

3. Download the dataset

Download WA_Fn-UseC_-Telco-Customer-Churn.csv from Kaggle and place it in the project root.

4. Update the script

In churn_prediction.py, replace the synthetic data section with:

df = pd.read_csv("WA_Fn-UseC_-Telco-Customer-Churn.csv")

5. Run

python churn_prediction.py

All plots will be saved to the output/ directory.

---

Models Used

Logistic Regression

A linear binary classifier that models churn probability using a sigmoid function. Simple, interpretable, and fast — but assumes linearity.

Random Forest

An ensemble of decision trees trained on random data subsets. Handles non-linear relationships well and is robust to overfitting.

Support Vector Machine (SVM)

Finds the optimal hyperplane to separate churn vs. non-churn customers. Uses an RBF kernel for non-linear classification.

---

Results

Model	Accuracy	Precision	Recall	F1-Score	ROC-AUC	Log-Loss
Logistic Regression	0.7864	0.6289	0.4867	0.5487	0.8148	0.4527
Random Forest	0.7693	0.5941	0.4282	0.4977	0.8007	0.4840
SVM	0.7779	0.6061	0.4787	0.5349	0.8048	0.4659

Key takeaways:

• Logistic Regression achieved the highest ROC-AUC (0.8148) and lowest log-loss
• SVM showed competitive performance with good class separation
• Random Forest had strong recall, identifying more actual churn cases
• All models struggled with the class imbalance (~73% non-churn vs ~27% churn)

---

Visualizations

Plot	Description
Churn Distribution	Class balance bar chart + churn by contract type
Box Plots	Feature distributions split by churn label
Correlation Heatmap	Pairwise correlations among numerical features
Confusion Matrices	Side-by-side for all three models
ROC Curve Comparison	AUC comparison across models
Learning Curves	Accuracy and log-loss vs. training size
Feature Importance	LR coefficients and RF Gini importance (top 15 features)

---

Future Improvements

• Handle class imbalance using SMOTE or class-weight adjustments
• Hyperparameter tuning with GridSearchCV or RandomizedSearchCV
• Ensemble methods — Voting Classifier or Stacking
• Deep learning — experiment with a simple neural network
• Model deployment — Flask/FastAPI REST API for real-time predictions
• Feature engineering — time-based features from tenure data

---

License

This project is for academic purposes only.