/cocktail_recommender

Primary LanguageJupyter Notebook

COCKTAIL RECOMMENDER PROJECT

I love craft cocktails, but during COVID I wasn't able to go out and enjoy them. During the lockdown, I decided to learn how to make my own cocktails. I bought a great cocktail book DEATH & CO and built up my home bar.

One thing I miss is being able to interact with a bartender and describe in vague terms what kind of cocktail I want. For example I can ask, "Do you have anything kinda smoky but spicy?" A good bartender can come up with a good cocktail to match. I wanted to train a model to do this for me.

BUILDING A MODEL TO MAP QUERIES TO COCKTAILS

For this project, I decided to build a model that uses NLP to rank cocktails that match a user's query.

STEPS:

  1. Get the data
  2. Extract ingredient categories for the cocktails
  3. Build the training set from Wikipedia and subreddit data
  4. Train the model and visualize results

GETTING THE DATA

The most important thing for me was to find a good source of creative cocktail recipes (no Cosmos messing up my dataset!). I found a great site upneat.rocks that had exactly what I was looking for. All cocktails from 3 famous craft cocktail books, including Death & Co!

I used Beautiful Soup (code in upneat_scraper.py) to extract all the contents

EXTRACTING INGREDIENTS

Ingredients had to be abstracted into categories for this to work. As seen in the example below, brand names are often found in the recipes (i.e. Siembra Azul Blanco Tequila) and I want this to simply be Tequila. This was somewhat arbitrary, but I managed to extract around 160 unique ingredients from the cocktail database.

GILDA COCKTAIL

  • 2 ounces Siembra Azul Blanco Tequila
  • .5 ounce Pineapple Juice
  • .5 ounce Lime Juice
  • .5 ounce Cinnamon Bark Syrup
  • Wheel of Lime (Garnish)

BUILDING A TRAINING SET FROM WIKIPEDIA AND REDDIT DATA

I selected the most common ingredients and scraped Wikipedia and subreddits (when they existed) corresponding to these ingredients (code in create_wiki_diciontaries.py and create_reddit_dictionaries.py)

Next, I separated all the ingredients into 5 general categories that cocktails tend to have:

  1. Spirits (Rum, Tequila, etc.)
  2. Juices (Lemon, Lime, etc.
  3. Liqueurs (Triple Sec, Kahlua, etc.)
  4. Syrups (Simple syrup, grenadine, etc.)
  5. Misc (Coffee, cream, etc.)

For each of these 5 categories, I built a model using Scikit-Learn's TFID Vectorizer to discover which words tended to be more frequent for a given ingredient (code in create_NLP_models.py)

DEMO (In TDI_Capstone notebook)

The model finds the most likely ingredients that match a query. Afterwards, cosine similarity is used to determine the closest cocktails to this query. I combine these results with simpler search methods that look for the query in the cocktail description and ingredients. The end result is the top N cocktails projected onto a 2d map that was derived with Multidimensional Scaling on the set of cocktail ingredients

Example Query: "smoky daiquiri"

JOVENCOURT DAIQUIRI

  • 2 ounces Barbancourt White Rum
  • .25 ounce Del Maguey Vida Mezcal
  • .75 ounce Lime Juice
  • .75 ounce Simple Syrup

Shake all the ingredients with ice, then strain into a coupe. No garnish.

PHIL WARD, 2007 This cocktail was my first epiphany about using mezcal in a mixed drink. Doing so is common now, but nobody was mixing with it back then.—PW

There is a method to get a score for each of the ingredients that were derived from the model: