Customer Churn Prediction in Telecom Industry

This project focuses on predicting customer churn in the telecom sector using machine learning techniques.
Customer churn refers to customers leaving a service provider, which directly impacts revenue.
The goal of this project is to identify customers who are likely to churn in advance, so that telecom companies can take proactive retention actions such as offering discounts or personalized plans.

The project is designed to be practical, interpretable, and deployable, rather than only accuracy-focused.

---

Problem Statement

In the telecom industry, acquiring new customers is significantly more expensive than retaining existing ones.
However, customer churn datasets are often imbalanced, and many machine learning models behave like black boxes, making business decisions hard to justify.

This project aims to:

• Predict churn accurately
• Handle class imbalance effectively
• Explain why a customer is predicted to churn
• Provide a simple web-based interface for real-time churn analysis

---

Dataset

• Dataset: Telco Customer Churn Dataset
• Source: Kaggle
• Records: 7,032 customers
• Target Variable: Churn (Yes / No)

Feature Types

• Demographic details (gender, senior citizen, dependents)
• Account information (tenure, contract type, payment method)
• Service usage (internet service, online security, streaming services)
• Billing details (monthly charges, total charges)

---

Project Workflow

1. Exploratory Data Analysis (EDA)

   • Churn distribution analysis
   • Feature-level churn trends
   • Identification of early churn patterns

2. Data Preprocessing

   • Handling missing values
   • Encoding categorical variables
   • Scaling numerical features
   • Class imbalance handling using SMOTE

3. Model Development

   • Logistic Regression (baseline)
   • Random Forest
   • XGBoost (final model)

4. Model Evaluation

   • Accuracy
   • Precision
   • Recall
   • F1-score
   • ROC–AUC
   • Confusion Matrix

5. Explainability

   • SHAP values to interpret model predictions
   • Identification of key churn drivers

6. Deployment

   • Streamlit-based web application
   • Real-time churn risk scoring
   • Downloadable churn-risk customer list

---

Models Used

Model	Purpose
Logistic Regression	Baseline comparison
Random Forest	Tree-based ensemble
XGBoost	Final optimized model

XGBoost was selected as the final model due to its:

• Strong performance on imbalanced data
• High recall and ROC-AUC
• Ability to capture non-linear feature interactions

---

Explainability (SHAP)

SHAP (SHapley Additive exPlanations) is used to explain model predictions.

Key insights from SHAP analysis:

• Contract type, tenure, and monthly charges are the strongest churn drivers
• Short-tenure, month-to-month customers show higher churn risk
• SHAP plots help translate model outputs into business-friendly insights

This makes the model suitable for real-world decision-making, not just prediction.

---

Streamlit Web Application

The trained and calibrated model is deployed using Streamlit.

Features:

• Interactive churn probability threshold selection
• Identification of churn-risk customers
• Ranked churn-risk list
• CSV download for retention campaigns

This allows business teams to use the model without coding knowledge.

---

Tech Stack

• Language: Python
• Libraries:
  • pandas, numpy
  • scikit-learn
  • XGBoost
  • SHAP
  • imbalanced-learn (SMOTE)
  • Streamlit
• Environment: VS Code, Jupyter Notebook

---

How to Run the Project

1. Clone the repository

git clone https://github.com/gokularaman-c/customer-churn-prediction.git
cd customer-churn-prediction

2. Install dependencies

pip install -r requirements.txt

3. Run the Streamlit app

streamlit run app.py

---

Key Outcomes

• Achieved strong recall and ROC-AUC for churn detection
• Built an interpretable ML pipeline using SHAP
• Developed a deployable, user-friendly churn prediction system
• Designed for both academic evaluation and industry use

---

Author

Gokularaman C

---