Recipe Recommender System

Driven by my curiousity of how Netflix, Youtube and Spotify serve personalized recommendations, I decided to learn how to create my own recommender system.

Machine Learning Problem: Given a person’s preferences in past recipes, could I predict other new recipes they might enjoy?

I created Seasonings, a Recipe Recommender System. The motivation behind this web app is to help users discover personalized and new recipes, and prepare for grocery runs! I received a lot early positive feedback and plan future improvements to the UX and model.

I had a lot of fun making this, and plan to use this whenever I need a jolt of inspiration in the kitchen!

Data

Data was scraped from allrecipes.com, as there was no public API. I narrowed the scope to focus on Chef John’s recipes (from FoodWishes.com).

Content Data

all_recipes.csv
1100+ Recipes from
460+ Cuisines & Categories

Content Data

all_users.csv
55K Users
73K Ratings

Tech Stack

Data Wrangling: pandas, numpy

Web Scraping: beautifulsoup, requests, regex

Please refer to web_scraper.py for more details

Model: scikit-learn, scipy

See requirements.txt for more

Web Framework: flask

Run app.py on localhost:5000 ```

Front End: html & css

Models

Please refer to model.py

Collaborative Filtering - Suggest recipes that other users similar to you also liked (Cosine Similarity)

If I liked Spaghetti Al Tonno, and another user similar to me liked Perfect Prime Rib and I haven’t tried it, the model would recommend that recipe.

Content Based Filtering - Suggest recipes that are similar to recipes that you like (Cosine Similiarity)

If I liked Spaghetti Al Tonno, the model would recommend Italian Meatballs, because Italian Meatballs are similar to Spaghetti, in terms of the categories both recipes share (Italian, World Cuisine).

Matrix Factorization - Suggest recipes that you like, uncover latent factors, in a lower dimensional space (Singular Value Decomposition)

If I liked Turkey, and I liked Cranberry Sauce, the model would recommend Pumpkin Pie because it picked up a latent factor that you liked Thanksgiving dishes, where the other models would not be able to.

Model Evaluation

My final model was a hybrid recommender that tackled the cold-start problem with a content recommender, augmented with user preferences, and factorization to rank recipes based on a voting classifier rule.

Screenshots

Onboarding

Results (Hybrid, Collaborative Filtering & Content Filtering)

References

Special thanks to Kim Falk’s book and also Maciej’s GitHub for reference during this journey.