ABOUT

Building a food recommender website

February, 2021 - March, 2021 Yonsei Big Data Conference Data Science Project

All our team members just love drinking beer and as we got tired of choosing the accompanying food, we wanted to ask the algorithm to do it for us.

Food Recommender System

Area of Focus

Collecting beer data and studying the pairing information, we processed our data. Then, we used the enormous recipe data to train and match the food options when the recommended food is inputted. The final results show the five most matching and suitable food items.

Website Deployment

Solution Design

Growing up mostly in Asia, us three were not really aqcuainted with the types of beer that are from elsewhere. Despite the ignorance, we love drinking beer, and decided to create our own food pairing algorithms while learning about some of them. We want to share our experience and provide pairing service to the public by launching a website.

https://www.homestratosphere.com/types-of-beer/

Beer Types

'Blonde Ale','Hefeweizen','Witbier','Classic Pilsner', 'British Style Beer','American Wheat Ale','India Pale Ale','Double/ Imperial IPA', 'Oktoberfest','Pale Ale','Red/ Amber Ale','Barley Wine','Brown Ale','Old/ Strong Ale', 'Abbey Dubbel','Dark Lager','Porter','Sweet/ Oatmeal Stout','Imperial Stout'

Data Type

1. Basic Beer Information
2. Food Pairing Instructions
3. 60,000 Recipe Data

1. Basic Beer Information

Data type 1 Text crawled from Craftbeer.com Data in the form of:

Beer Name
Original Gravity (OG)
Final Gravity (OG)
Alcohol By Volume (ABV)
International Bitterness Units (IBUs)
Bitterness Ratio (BU:GU)
Volumes of CO2 (v/v)
Apparent Attenuation (AA)

2. Food Pairing Instructions

Data type 2 Understand basic guidelines to match the food and beer; commonly paired tastes and types of food provided. Collected from Brewers Association, Shari's Berries, Hangout Restaurant

Completed data are:

19 kinds of beer
84 ingredient lists
44 desserts
19 spices
21 whole foods

3. Recipe Data

Data type 3 60,000 Recipe Data from Kaggle What's Cooking?

Logical conclusion that machine learning the recipe ingredients per recipe could obtain the function that recommends well-pairing ingredient from another.
"id", "cuisine", "ingredients" in · Calculate co-occurrence
"id", "cuisine", "ingredients" in .json
Predict and frame the realtionship by checking the level of co-occurrence
Zip the ingredient file sorted by occurrence

Machine Learning

Modeling Work Flow

Based on the created data, we were able to set a function where we simply print the ingredients in an ascending order after inputing beer name.
Using the co-occurrence data, other recommeended ingredients are simultaneously printed.
We then finalized the system by calculaing the PMI and omitted those are very low or is negative value. PMI(A,B) = logP(A,B) / (P(A)*P(B))
We finally are able to print the ingredients similar to that of the main ingredients (n=84) suited for each beer.

Website

Although deploying both Django and Node.js is practically impossible, we used two seaparate servers for this case and decided to unite to one of them in future projects.
Using cytoscape.js, a mindmap tool based on Node.js, we could visualize the 20 ingredients that best matched the each of the main ingredients paired with the beer, and that prints the results.
The size of the first node is proprotional to the pairing score with the berr, while the second node is proportional to each ingredient's pairing score with the main ingredient.
When the user hovers over the desired ingredient, the UI is set to emphasize its name and color coded the ingredients hierarchically.

Future Works

Adding more beer or types of alcohol
Recommender system according to the user's preference
Uniting its backend framework
Launching the website/app

Demonstration

Pairing Visualization

A Supplementary Individual Work by Eunsol Choi

Paired two kinds of beer according to each month and formed a calendar design.

⇽Previous Project
Apartment Prediction

Next Project⇾
NLP Chicken