twotower-yelp-baseline

Two-Tower Recommendation Model (Yelp Baseline)

This repository contains a baseline Two-Tower recommendation model trained on the Yelp Dataset.

📘 Model Summary

• Architecture: Two-Tower (User Tower + Item Tower)
• User Tower: Embedding table trained end-to-end
• Item Tower: MLP taking pretrained POI embeddings (772-dim)
• Loss: Binary cross entropy with negative sampling
• Negative Samples per interaction: 10
• Training Epochs: 3 (CPU-based)

The model was trained as a foundational retrieval baseline to validate end-to-end recommendation pipeline feasibility.

📂 Included Files

• twotower_model.pt — Trained model weights
• .gitattributes — Git LFS tracking file

⚠ Limitations

This is a baseline model and achieves low performance due to:

• user_id-only embeddings (no behavioral or text features)
• CPU-only training (limited epochs)
• sparse interactions across many users

🚀 Future Improvements

• Add user textual embeddings (LLM-based persona summaries)
• Add sequence-based behavioral embeddings
• Apply hard negative sampling & better contrastive losses
• Train on GPU for larger batch sizes & deeper models

---

📄 Usage Example (Python)

import torch
from model import TwoTowerModel  # Your model class

model = TwoTowerModel(num_users=851921, emb_dim=128)
model.load_state_dict(torch.load("twotower_model.pt"))
model.eval()