📊 End-to-End ML Deployment: Telco Customer Churn Project

🌐 Live Demo

Click 👉

---

📖 Project Overview

Customer churn is a major challenge for telecom companies — retaining customers is often more cost-effective than acquiring new ones. This project builds an end-to-end machine learning pipeline to predict churn, explain drivers of churn, and segment customers into actionable groups for better retention strategies.

The project includes:

• EDA → Explore churn patterns, tenure, contracts, charges.
• Customer Segmentation → KMeans (baseline) vs HDBSCAN (tuned).
• Churn Prediction → Logistic Regression baseline vs advanced ensemble models (Random Forest, XGBoost, Voting Classifier).
• Explainability → SHAP summary & waterfall plots.
• Interactive App → Built with Streamlit, deployed on Streamlit Cloud.

---

📂 Dataset

Dataset: WA_Fn-UseC_-Telco-Customer-Churn.csv

• Target: "Churn" (Yes/No)

• Features include:

  • Demographics → Gender, Senior Citizen, Dependents, Partner
  • Services → Phone, Internet, Tech Support, Streaming, Security
  • Account Info → Tenure, Contract, Billing, Payment Method
  • Charges → Monthly & Total Charges

---

🧪 Methods & Models

🔎 Exploratory Data Analysis (EDA)

• Customers with fiber optic internet churn the most (pricing/service quality issues).
• DSL customers churn less, possibly due to stable pricing or loyalty.
• High churn in first 5 months → critical onboarding phase.
• Long-tenure customers (>24 months) show significantly lower churn rates.

👥 Customer Segmentation (Unsupervised Learning)

• Cluster 0 — Budget Loyalists → Minimal services, mailed check payments, stable.
• Cluster 1 — At-Risk Premiums → Fiber optic, month-to-month, electronic check, highest churn risk.
• Cluster 2 — Balanced Mainstream → Moderate DSL usage, mixed services, mid-spenders.
• Cluster -1 — Drifters → DSL, no phone, low commitment.

📊 Churn Prediction Models

• Logistic Regression (baseline)
• Random Forest (ensemble)
• XGBoost (boosted trees)
• Voting Classifier (combined)

🔍 Explainability (SHAP)

• Feature importance ranking.
• SHAP summary plots + waterfall plots for individual predictions.

---

🚀 Deployment

• Streamlit App for interactive visualization and prediction.
• Dockerized for reproducibility.
• Deployed on Streamlit Cloud with a public link.

---

⚙️ Installation & Usage

1. Clone Repo

git clone https://github.com/<your-username>/Customer_Churn_Prediction.git
cd Customer_Churn_Prediction

2. Install Requirements

pip install -r requirements.txt

3. Run Locally

streamlit run scripts/app.py

App runs at: http://localhost:8501

4. Run with Docker

docker build -t churn-app .
docker run -p 8501:8501 churn-app

---

📦 Project Structure

Customer_Churn_Prediction/
│
├── data/                 # feature store JSON (not raw data)
├── models/               # saved ML models (.joblib)
├── reports_app/          # plots & visualizations
├── scripts/              # Streamlit app (app.py) & utilities
├── src/                  # preprocessing, feature engineering, utils
├── config.json           # config settings
├── requirements.txt      # dependencies
├── Dockerfile            # container setup
└── README.md             # this file

---

🛠️ Tech Stack

• Python: pandas, numpy, scikit-learn, xgboost, shap, hdbscan, umap
• Visualization: matplotlib, seaborn, streamlit
• MLOps Tools: Docker, GitHub, MLflow (Experimental Tracking)
• Deployment: Streamlit Cloud

---

📌 Next Steps

• Extend segmentation with deep embeddings.
• Add hyperparameter search with Optuna.
• Deploy with a custom domain using Render or Railway.

---

👤 Author

Developed by Tszon Tseng

• 💼 Passionate about Data Science & AI
• 🚀 Building end-to-end ML pipelines
• 🌐 LinkedIn Profile

---

✨ With this app, telecom providers can predict churn, understand why customers leave, and design better retention strategies.