Universal Recommender

A recommendation engine built on a domain-agnostic architecture, demoed and deployed on MovieLens-1M (1M movie ratings, 6,040 users, 3,706 movies).

Live app: https://rec-sys-gfrqwtuewmgqeuapwosliv.streamlit.app/ Repo: https://github.com/Aryan8912/Rec-Sys

What it does

Enter a User ID, pick a recommendation strategy, get a ranked list of movie recommendations — optionally with an AI-generated explanation for the top pick.

Two ways to get recommendations:

• Single model: choose directly from Popularity, Item-Item CF, Matrix Factorization, Two-Tower, or Neural CF
• Full funnel: Two-Tower generates a user embedding → FAISS retrieves ~200 candidates → Neural CF re-ranks them to the final list (the same retrieval-then-ranking pattern used in production systems like YouTube)

Setup (local)

python3 -m venv .venv
source .venv/bin/activate        # .venv\Scripts\activate on Windows
pip install -r requirements.txt

Train all models and save artifacts:

python src/train.py

Run the app:

streamlit run app/streamlit_app.py

Project structure

├── data/                  MovieLens-1M CSVs (ratings, movies, users)
├── src/
│   ├── data_pipeline.py   loads and splits data
│   ├── models/            Popularity, Item-CF, MF, Two-Tower, Neural CF
│   ├── retrieval/         FAISS candidate retrieval
│   ├── rag/                semantic embeddings + LLM explanation generator
│   ├── evaluation/        Precision@K, Recall@K, NDCG@K, Coverage
│   └── train.py           trains everything, saves to artifacts/
├── app/
│   └── streamlit_app.py   the deployed UI
├── artifacts/             trained models (committed — app loads these,
│                            does not retrain on startup)
├── docs/
│   ├── DOCUMENTATION.md   full write-up: architecture, methodology,
│   │                        evaluation, test cases, limitations
│   └── COMPARISON.md      comparison against YouTube's public architecture
└── requirements.txt

Results

Model	Precision@10	Recall@10	NDCG@10	Coverage
Popularity	0.0668	0.0427	0.0762	0.0232
Item-Item CF	0.0838	0.0770	0.1024	0.3030
Matrix Factorization	0.0772	0.0575	0.0897	0.2372
Two-Tower	0.0670	0.0420	0.0756	0.1082
Neural CF	0.0702	0.0440	0.0781	0.1082

Full methodology, architecture diagrams, and honest discussion of why the deep models don't outperform Item-CF on this dataset are in DOCUMENTATION.md.