Feature Store with Feast

Centralized feature engineering and serving platform built on Feast for healthcare fraud detection. This project implements a complete feature store pipeline with offline (file-based) and online (Redis) stores, supporting both batch training and low-latency real-time inference.

Architecture

Raw Claims Data (Parquet)
        |
  Feature Pipeline
        |
   +-----------+
   |   Feast   |
   +-----+-----+
   |           |
Offline Store  Online Store
 (Parquet)      (Redis)
   |               |
Training Data   Real-time
Retrieval       Serving

Features

• Entity Definitions: Provider, Patient, and Claim entities with proper join keys
• Feature Views: Provider-level, patient-level, and claim-level feature aggregations
• On-Demand Features: Computed features like amount deviation from provider averages
• Feature Service: Unified feature retrieval interface for fraud detection models
• Materialization Pipeline: Scheduled push from offline to online store
• REST API: FastAPI-based feature serving endpoints
• Synthetic Data Generation: Realistic healthcare claims data with configurable fraud rates

Project Structure

feast-feature-store/
├── feature_repo/
│   ├── feature_store.yaml      # Feast configuration
│   ├── entities.py             # Entity definitions
│   ├── feature_definitions.py  # Feature view definitions
│   └── data/                   # Registry and offline store
├── src/
│   ├── data_generator.py       # Synthetic data generation
│   ├── feature_pipeline.py     # Feature computation pipeline
│   ├── model_training.py       # ML training with feature store
│   └── api.py                  # REST API for feature serving
├── configs/
│   └── config.yaml             # Pipeline configuration
├── data/                       # Generated parquet files
├── requirements.txt
└── README.md

Quick Start

Prerequisites

• Python 3.9+
• Redis server running on localhost:6379
• pip package manager

Installation

pip install -r requirements.txt

Generate Synthetic Data

python src/data_generator.py

Apply Feature Definitions

cd feature_repo
feast apply

Run Feature Pipeline

python src/feature_pipeline.py

Materialize Features to Online Store

feast materialize 2023-01-01T00:00:00 2024-01-01T00:00:00

Train Model

python src/model_training.py

Start API Server

uvicorn src.api:app --host 0.0.0.0 --port 8000

Feature Definitions

Provider Features

Feature	Description
billing_frequency	Number of claims submitted per month
avg_claim_amount	Average dollar amount per claim
specialty	Provider medical specialty code
patient_count	Number of unique patients seen
claim_denial_rate	Fraction of claims denied

Patient Features

Feature	Description
age	Patient age in years
num_claims	Total number of claims filed
total_billed	Cumulative billed amount
avg_days_between_claims	Average gap between consecutive claims
num_unique_providers	Number of distinct providers visited

Claim Features

Feature	Description
amount	Claim dollar amount
procedure_code	CPT procedure code
diagnosis_code	ICD-10 diagnosis code
days_since_last_claim	Days since patient's previous claim
provider_patient_pair_frequency	How often this provider-patient pair appears

API Endpoints

Method	Endpoint	Description
POST	/features/online	Retrieve online features for an entity
POST	/features/historical	Retrieve historical features with timestamps
POST	/materialize	Trigger feature materialization
GET	/feature-views	List all registered feature views
GET	/health	Health check

Configuration

All pipeline parameters are configured via configs/config.yaml, including data generation settings, materialization windows, and feature retrieval timeouts.

License

MIT License