git clone https://github.com/featurestorebook/mlfs-book.git
cd mlfs-book
cp .env-env .env

Machine Learning Feature Store Book - Example Projects

This repository contains three complete machine learning systems demonstrating different architectural patterns and use cases. Each project showcases best practices for building production ML systems with feature stores.

Quick Start

git clone https://github.com/featurestorebook/mlfs-book.git
cd mlfs-book
cp .env-env .env

1. Create an account on https://run.hopsworks.ai
2. (Create an API key on Hopsworks)[https://docs.hopsworks.ai/latest/user_guides/projects/api_key/create_api_key/] and set its value in the HOPSWORKS_API_KEY line in the .env file
3. Update HOPSWORKS_PROJECT in the .env file to the name of your project

To run the starter titanic batch ML system:

cd titanic
source setup-env.sh
inv all

To run the air quality batch ML system:

cd airquality
source setup-env.sh
inv all

To run the credit-card fraud real-time ML system:

cd ccfraud
source setup-env.sh
inv all

Note: This requires a command line on your laptop and has been tested on Linux, Mac, and Windows Subsystem for Linux (WSL).

Projects

1. Batch Air Quality Predictions

A batch ML system that forecasts air quality (PM 2.5) for the next 7 days using real-time weather data and historical air quality measurements from public sensors (https://waqi.info/). The system uses XGBoost to generate predictions and displays them on an interactive dashboard.

Type: Batch Predictions Model: XGBoost Regression Features: Weather forecasts, historical air quality data Output: 7-day air quality forecast dashboard

Optional LLM Enhancement: The air quality service can be augmented with LLM capabilities for natural language queries about air quality forecasts and historical trends.

View Project →

2. Real-Time Credit Card Fraud Detection

A real-time ML system that detects fraudulent credit card transactions using streaming data and real-time feature engineering. This project demonstrates how to build low-latency prediction services that process transactions as they occur, using Feldera for stream processing.

Type: Real-Time Predictions Model: Binary Classification (XGBoost) Features: Real-time transaction aggregates, velocity features Output: Fraud probability for each transaction Data: Synthetic credit card transactions

View Project →

3. Starter Titanic Batch Predictions

A starter ML system for predicting Titanic passenger survival. This project provides a simple, well-structured example of a batch ML pipeline using the classic Titanic dataset for training and synthetic data for generating daily predictions. Includes both a batch dashboard and an interactive UI.

Type: Batch Predictions Model: Binary Classification (XGBoost) Features: Passenger demographics, ticket class, family size Output: Dashboard and interactive Gradio UI Data: Historical Titanic dataset + synthetic passengers

View Project →

Architecture

Each project follows a similar architectural pattern:

• Feature Store: Central repository for feature storage and retrieval
• Feature Pipeline: ETL jobs that compute and store features
• Training Pipeline: Model training with versioning and experiment tracking
• Inference Pipeline: Batch or real-time predictions
• UI/Dashboard: Visualization of predictions and monitoring

Getting Started

Each project can be run independently. Navigate to the project directory and follow the instructions in its README.md file.

Additional Resources

You can find additional tutorial instructions in this Google Doc.