Recommender Systems on Food.com Rating

Overview

This project develops advanced recommender models to improve personalized recipe recommendations on Food.com. The models utilize user reviews and ratings to predict user satisfaction and enhance the browsing experience.

Project Summary

The project utilizes three main models to predict recipe ratings:

1. Baseline Linear Regression Model - Establishes fundamental user and item biases.
2. Sentiment Analysis Model - Incorporates sentiment from user reviews to capture emotional feedback.
3. SVD++ Model - Uses matrix factorization to uncover deeper, implicit user preferences.

Combining sentiment analysis with SVD++ in an ensemble model, we achieved our best accuracy with an MSE of 0.234, significantly enhancing prediction accuracy.

Datasets

1. Raw Recipes Dataset - Contains 230k+ recipes with details such as cooking time, ingredients, nutrition, etc.
2. Raw Interactions Dataset - Includes 1.1 million user interactions (ratings, reviews).

The datasets were merged and cleaned, resulting in 650k entries optimized for predictive modeling.

Models and Results

• Baseline Linear Regression - Achieved an MSE of 0.476.
• Sentiment Analysis Model - Improved prediction accuracy, achieving an MSE of 0.341.
• SVD++ Model - Leveraged implicit interactions, reaching an MSE of 0.328.
• Ensemble Model - Combined Sentiment Analysis and SVD++ for the best performance with an MSE of 0.234.

Conclusion

This ensemble approach effectively combines quantitative and qualitative data, providing enhanced personalized recipe recommendations and a more satisfying user experience on Food.com.