Amazon Electronics Recommender System

This project is an end-to-end recommendation system built from the Amazon Reviews 2023 Electronics dataset. It covers the full lifecycle:
data preparation → model development → evaluation → scalable deployment as a REST API on Azure Container Apps.

Live Demo (Azure Container App) (Note: The app scales to zero when idle; the first request may take up to half a minute to start, but subsequent queries respond instantly.)

---

Project Highlights

• Built a scalable recommendation system on Amazon Reviews dataset (41K users, 137K items, 391K interactions).
• Improved ranking performance with metadata + SBERT text embeddings.
• Designed a two-stage retrieval pipeline (FAISS + MLP rerank) for sub-100ms inference.
• Deployed as a REST API on Azure Container Apps with Docker and CI/CD workflows.

---

Key Features

• Data Engineering

  • Started from a strict 5-core subset (users and items with ≥5 interactions).
  • Sampled ~2.5% of users, keeping all their interactions to preserve temporal coherence.
  • Retained items with at least one interaction among these users.
  • Final dataset: 41,909 users, 136,934 items, 390,757 interactions (sparser than full 5-core).
  • Applied chronological train/val/test splits and engineered features (price, ratings, categories, text embeddings).

• Modeling

  • Implemented and compared four model families:
    • Matrix Factorization (baseline CF)
    • Hybrid CF + metadata
    • Hybrid CF + SBERT text embeddings
    • Two-Tower MLP with metadata + text (best performing)
  • Optimized with Bayesian Personalized Ranking (BPR) loss and multi-seed validation.
  • Achieved consistent improvements in NDCG@10 with text + MLP models.

• Retrieval & Serving

  • Built a two-stage pipeline: FAISS retrieval for candidate generation + MLP re-ranking.
  • Packaged inference assets (model, FAISS index, metadata, embeddings) for reproducible deployment.
  • Exposed as a REST API with FastAPI, returning JSON responses for easy integration.
    • /health → system status
    • /users/sample → sample user IDs
    • /recommend/{user_id} → personalized recommendations
    • /recommendation/demo → demo with random user

• Deployment

  • Containerized with Docker (python:3.12-slim base).
  • Deployed as a public Azure Container App with scale-to-zero for cost efficiency.
  • Inference latency <100ms per request on CPU.

---

Tech Stack

• Languages & Libraries: Python, PyTorch, FAISS, FastAPI, NumPy, Pandas
• NLP: SBERT (SentenceTransformer all-MiniLM-L6-v2) for item title embeddings
• Experimentation: Hyperparameter search, multi-seed evaluation, HR@K & NDCG@K metrics
• Deployment: Docker, Azure Container Apps

---

Reproducibility

1. Train model:

   python scripts/Recommendation_System/train.py

2. Build FAISS index:

   python scripts/Recommendation_System/build_FAISS_index.py

3. Run API locally:

   uvicorn scripts.Recommendation_System.recommender_api:app --reload

4. Build Docker image:

   docker build -t recsys-api -f DOCKERFILE .
   docker run -p 8000:8000 recsys-api

Docker Hub Image

The API is also available as a prebuilt Docker image:

Docker Hub – rahulk98/amazon-electronics-recommender-api

Run locally with:

docker pull rahulk98/amazon-electronics-recommender-api:latest
docker run -p 8000:8000 rahulk98/amazon-electronics-recommender-api:latest

---

Evaluation Results (Multi-Seed)

Model	NDCG@10 (mean)	Notes
Collaborative Filtering	0.3051	Baseline CF (BPR loss)
Hybrid (metadata)	0.2861	Used category + numeric only
Hybrid + Text	0.3311	SBERT title embeddings added
Two-Tower MLP	0.3342	Best overall, scalable

Improved NDCG@10 from 0.305 (CF baseline) to 0.334 (Two-Tower MLP + text), confirming the value of combining collaborative, metadata, and text signals.

---

API Demo Screenshots

Demo Recommendations

Sample recommendations returned for a random user via /recommendation/demo.

Swagger Docs

Interactive API documentation generated by FastAPI.

Single User Query

Personalized recommendations for a specific user via /recommend/{user_id}.