ReciSys: Hybrid Recipe Recommendation with Attention & NLP

This project studies rating prediction on the Food.com dataset using modern recommender system techniques.

I explore two complementary predictive tasks:

1. Hybrid Recommendation (Collaborative + Content-Based)

   • Predict user–recipe ratings using an attention-based neural recommender
   • Combines user embeddings, item embeddings, and rich recipe metadata
   • Model: SideFeatAttAE

2. NLP-Based Rating Prediction

   • Predict ratings directly from review text
   • Models: Fine-tuned DeBERTa and TF-IDF + Ridge regression

Together, these tasks examine how structured interactions and natural language capture user preferences from different perspectives.

Dataset: Food.com (Majumder et al., EMNLP 2019)

• 230K+ recipes
• 1M+ user interactions
• 18 years of data (2000–2018)

After filtering:

• ~180K recipes
• ~700K interactions

Original dataset goal: personalized recipe generation
This project focuses on rating prediction and representation learning.

Goal: Predict a user's rating for a recipe.

Inputs

• User ID
• Recipe ID
• Recipe side features:
  • cooking time
  • number of steps
  • number of ingredients
  • nutrition vector
  • recipe tags

Model

• User & item embeddings
• Side-feature embedding via MLP
• Multi-head attention over [user, item, side] tokens
• MLP head + bias terms

Baselines

• Global mean
• User mean
• Item mean
• Matrix Factorization (SVD)

Metric

• Mean Squared Error (MSE)

Goal: Predict rating from review text alone.

Models

• Fine-tuned DeBERTa-v3 (regression)
• TF-IDF + Ridge regression

Why NLP?

• Captures sentiment, negation, and subtle phrasing
• Complements collaborative signals

Metric

• Mean Squared Error (MSE)

Model	Test MSE
SideFeatAttAE	0.712
Matrix Factorization (SVD)	0.753
Global Mean	0.845
TF-IDF + Ridge (NLP)	0.378
DeBERTa (NLP)	0.286

• Attention-based hybrid models outperform classical CF baselines

• Rich side information improves robustness and cold-start behavior

• NLP models can recover ratings extremely well from language alone

• The architecture generalizes beyond Food.com to many recommendation domains

• notebook/ReciSys.ipynb – full implementation and experiments

• slides/ReciSys.pptx – presentation walkthrough

• report/ReciSys_Report.pdf – written report

• watch: video presentation

Dataset Citation

This project uses the Food.com dataset introduced in:

Bodhisattwa Prasad Majumder*, Shuyang Li*, Jianmo Ni, Julian McAuley.
Generating Personalized Recipes from Historical User Preferences.
EMNLP–IJCNLP, 2019.

• Paper: https://cseweb.ucsd.edu/~jmcauley/pdfs/emnlp19c.pdf
• Dataset: https://www.kaggle.com/datasets/shuyangli94/food-com-recipes-and-user-interactions

If you use this dataset, please cite the paper above.