Churn Prediction with Cat-Boost: A End-to-End Project for Telco Customers

Predicting customer churn with machine learning isn’t just a portfolio project — it’s a business-critical use case. In this post, I’ll walk you through how I built a production-grade, end-to-end churn prediction system using Cat-Boost, Stream-lit, Docker, and more.

Problem Statement: Why Does Churn Matter?

Churn — when a customer leaves a service — is one of the biggest threats to subscription-based businesses like telecom providers. Acquiring a new customer is 5–10x more expensive than retaining an existing one. My goal? Build a system that predicts which customers are likely to churn, so the business can proactively intervene and retain them.

Tools & Tech Stack

• Data Analysis: Pandas, Matplotlib, Seaborn

• Feature Engineering: Scikit-learn, Category Encoders

• Model Training: CatBoost, XGBoost, Random Forest, Logistic Regression, SVM

• Evaluation: Precision, Recall, F1, AUC

• Deployment: Stream-lit, Docker

• Code Structure: Production-grade with config mgmt, unit tests

Project Architecture:

Raw Telco Data → EDA & Cleaning → Feature Engineering → Train 5 ML Models → Model Selection (Cat-Boost wins) → Stream-lit Dashboard → Dockerized Deployment

EDA: Understanding the Churn Story

The Telco dataset included features like:

• Contract, InternetService, PaymentMethod

• tenure, MonthlyCharges, TotalCharges

• Churn (target)

Key insights:

• Short-tenure customers churn more

• Month-to-month contracts have high churn

• Fiber optic customers and those with electronic checks were more likely to churn

Feature Engineering: Turning Raw Data into Gold Steps included:

• Handling missing values (TotalCharges)

• Label encoding binary features

• One-hot encoding categorical features

• Scaling numerical values

• Creating interaction terms (e.g., charges_per_month = TotalCharges / tenure)

Model Training & Evaluation

I trained and compared 5 models:

1. CatBoost

2. XGBoost

3. Random Forest

4. Logistic Regression

5. SVM

Evaluation Metrics:

• Accuracy: not enough (imbalanced data)

• Used F1 Score, Precision, Recall, and AUC-ROC

CatBoost outperformed all others:

• F1: 0.83

• AUC: 0.89

• No need for heavy encoding = simpler pipeline

Model Explainability I used:

• Feature importance from CatBoost

• SHAP values to visualize individual predictions

Most important features:

• Contract

• tenure

• MonthlyCharges

• InternetService

Streamlit Dashboard Built an interactive web app using Streamlit. Key features:

• Upload a new dataset

• Visualize churn distribution

• Predict churn for individual users

• Explain predictions using SHAP plots

Dockerized Deployment Wrapped the entire app in a Docker container:

• Reproducible across machines

• Easily deployable to Render / HuggingFace / Streamlit Cloud

Also included:

• .env for secrets

• config.yaml for tunables

• pytest tests for pipelines

Key Takeaways

• Hands-on with real ML tools, not just theory

• Built a real product, not just a notebook

• Learned how to deploy ML apps for the real world

• Streamlit + Docker is a game-changer for ML portfolios

Project Links

• GitHub Repo: https://github.com/SANJAYRAM-DS/churn-ml-data-to-docker

• Live Demo: https://churn-ml-data-to-docker-kvrk8rh44zucasc3vcmx6n.streamlit.app

• Let's connect on LinkedIn: www.linkedin.com/in/sanjayram-data

Final Thoughts This project took me from data wrangling to full-stack ML deployment. I now understand what end-to-end machine learning really means. Hope this inspires you to go beyond Jupyter notebooks and build real, deployable ML systems!

Let me know your thoughts — and feel free to fork the repo to try it yourself