Telecom Customer Churn Prediction: From Data Analysis to Production-Grade MLOps Practice

This project is an end-to-end machine learning solution addressing the pain point of customer churn in the telecom industry. It predicts customer churn risk using gradient boosting models and integrates a complete MLOps pipeline with Streamlit interactive dashboards, DVC data version control, MLflow experiment tracking, and Kubernetes containerized deployment, achieving a closed loop from data exploration to production-grade deployment.

Customer churn is a costly operational pain point in the telecom industry—acquiring new customers costs 5-7 times more than retaining existing ones. Traditional post-hoc remediation strategies have limited effectiveness; proactively identifying at-risk customers and intervening early is key. The dataset for this project covers multi-dimensional information of 7043 customers (demographics, service subscriptions, billing, contract types, etc.), providing a foundation for model analysis.

Data Processing and Feature Engineering

Raw data is cleaned (e.g., type conversion and missing value handling for the TotalCharges column) and several features are derived: tenure_group (tenure grouping), num_services (number of services), is_longterm (long-term contract flag), has_support (technical support subscription), charges_per_month (average monthly charges), is_high_value (high-value customer flag), etc.

Model Selection and Performance

The gradient boosting machine (GBM) is used to adapt to tabular data, and SMOTE is applied to address class imbalance. Model performance on the test set: accuracy ≥88%, AUC-ROC ≥0.85, recall rate for churned customers ≥71%, precision ≥76%, which can support effective intervention strategies.

Customers are divided into 4 groups using K-Means clustering:

1. Loyal Long-term Users: Long tenure, annual contracts, low churn risk—suggest upselling premium services;
2. New High-Spending Users: Short tenure, high monthly spending, extremely high churn risk—need exclusive initial offers and VIP services;
3. Economical Monthly Users: Low monthly spending, pay-as-you-go, medium churn risk—suggest contract upgrade incentives;
4. Stable Mid-Tier Users: Medium tenure, multiple service subscriptions, low churn risk—suggest cross-selling support service packages.

Data and Model Version Control

Using DVC + DagsHub to implement data/model versioning to ensure experiment reproducibility.

Experiment Tracking

MLflow records experiment parameters and metrics, and registers the best model.

Containerization and Deployment

Docker packages the application to ensure environment consistency; Kubernetes orchestration achieves high availability (auto-scaling, self-healing); GitHub Actions CI/CD automates the process: code push → data pull → training → validation → Docker build → K8s update.

The Streamlit dashboard includes five modules:

• Overview Panel: Displays churn rate, contract distribution, and revenue impact;
• EDA Module: Interactive filtering and feature distribution charts;
• Churn Predictor: Inputs customer information and returns risk scores and driving factors;
• Segmentation Visualization: PCA dimensionality reduction to show cluster distribution;
• Revenue Simulator: Simulates the revenue impact of retention strategies. It has been deployed to Streamlit Cloud for direct use by business users.

Key Findings (SHAP Analysis)

1. Contract type: Monthly contract customers have 3x higher churn rate than annual contract customers;
2. Tenure: Churn risk is highest in the first 12 months of service;
3. Monthly spending: Customers with high spending but low perceived value are prone to churn.

Business Recommendations

• Promote long-term contracts and incentivize monthly contract users to upgrade;
• Design a "new user care" program to reach new users at key touchpoints;
• Provide personalized services for high-spending users. It is expected to reduce overall churn rate by 10-15% and increase high-value customer retention rate by 20%+.