Workflow-CI: MLflow Customer Churn Prediction Pipeline

A CI/CD pipeline that trains a Random Forest classifier for customer churn prediction using MLflow, containerizes the model with Docker, and automatically deploys it to Docker Hub.

Overview

This project implements an automated machine learning workflow that:

1. Trains a model using scikit-learn's Random Forest Classifier on the Telco Customer Churn dataset
2. Tracks experiments with MLflow for reproducibility and model versioning
3. Builds a Docker image containing the trained model
4. Pushes the container to Docker Hub for deployment

Project Structure

Workflow-CI/
├── .github/
│   └── workflows/
│       └── main.yml          # GitHub Actions CI/CD workflow
├── MLProject/
│   ├── MLProject             # MLflow project configuration
│   ├── conda.yaml            # Conda environment specification
│   ├── modelling.py          # Model training script
│   ├── DockerHub.txt         # Docker Hub repository URL
│   └── WA_Fn-UseC_-Telco-Customer-Churn_preprocessing.csv  # Preprocessed dataset
├── mlruns/                   # MLflow tracking directory
└── README.md

Dataset

The project uses the Telco Customer Churn dataset, which has been preprocessed with:

• Numerical feature scaling (standardization)
• Categorical feature encoding (one-hot encoding)
• Tenure binning

Target variable: Churn (binary classification - 0 or 1)

Model Configuration

The Random Forest classifier is configured with the following default parameters:

Parameter	Default Value
n_estimators	505
max_depth	37

These parameters can be customized when running the MLflow project.

Dependencies

The project uses the following main dependencies:

• Python 3.13.9
• MLflow 2.18.0rc0 (as specified in conda.yaml) and MLflow 2.22.0 (as used in CI/CD workflow)
• scikit-learn 1.7.2
• pandas 2.3.3
• numpy 2.3.5

See MLProject/conda.yaml for the complete dependency list.

Running Locally

Prerequisites

• Python 3.12+
• MLflow installed (pip install mlflow)
• Conda (optional, for environment management)

Run the MLflow Project

# Using local environment
mlflow run ./MLProject --env-manager=local

# Using conda environment
mlflow run ./MLProject

# With custom parameters
mlflow run ./MLProject -P n_estimators=100 -P max_depth=20 --env-manager=local

CI/CD Pipeline

The GitHub Actions workflow (.github/workflows/main.yml) automates the entire pipeline:

Triggers

• Push to main branch
• Pull requests to main branch
• Manual trigger via workflow_dispatch

Pipeline Steps

1. Checkout - Clone the repository
2. Setup Python - Install Python 3.12.7 (CI/CD) or Python 3.13.9 (local conda environment)
3. Install Dependencies - Install MLflow
4. Run Modelling - Execute the MLflow project
5. Build Docker Image - Create container with mlflow models build-docker
6. Push to Docker Hub - Deploy the containerized model

Required Secrets

Configure these secrets in your GitHub repository settings:

Secret	Description
DOCKER_HUB_USERNAME	Docker Hub username
DOCKER_HUB_ACCESS_TOKEN	Docker Hub access token
USERNAME	Git username for commits
EMAIL	Git email for commits

Docker Deployment

The trained model is available as a Docker image:

# Pull the latest model
docker pull pahlawandocker/churn:latest

# Run the model server
docker run -p 5000:8080 pahlawandocker/churn:latest

The model server exposes a REST API for predictions at http://localhost:5000/invocations.

MLflow Tracking

Experiment runs are tracked in the mlruns/ directory. Each run includes:

• Model artifacts
• Accuracy metrics
• Input examples
• Model signature

License

This project is open source and available for educational and research purposes.