Content-based model proof of concept

Demo

Note how the model updates online because of how fast the underlying matrix operations are!

This recommender will take user preferences for recipes and use these preferences to find the user's preferences for ingredients. Using the ingredients, we are then able to search among all recipes which have a similar ingredient matrix to that which the user prefers.

We will use TF-IDF as the measure of similarity between recipes. 83% of text-based recommenders use TF-IDF.

Getting Started

Install python 3.6.3:

pyenv install 3.6.3

Create content_poc virtual env:

pyenv virtualenv 3.6.3 content_poc

Install requirements:

pip install -r requirements.txt

Run the main file. You will be asked for a user id; feel free to give it any number you like (as long as you can remember it):

python content_based_meal_recommendation_main.py

You will be given 3 recipes to rate as a cold-start. Accepted inputs are 1,0,-1.

You will then be given a list of recipes that match your preferences best. If you enter your user id again, you can choose to rate more recipes for more accurate results or just retrieve predictions.

Requirements

Python 3.6.3
Pandas
scikit-learn

Dataset

train.json is a dataset consisting of a list of recipes where each recipe has a list of ingredients as well as an associated cuisine. We do not use the cuisine as a factor in training although it's useful in proving that our predictions are accurate.

TF-IDF

TF stands for Term Frequency which is the number of times a word/term appears in our dataset.

IDF stands for Inverse Document Frequency which is the inverse of the number of documents in which a word appears.

TF-IDF increases proportionally to the number of times a keyword appears but is offset by the number of couments in which it appears.

TF-IDF = TF * IDF

Creating user profiles

We have:

A list of recipes
User ratings for each recipe

For each ingredient in a recipe, we calculate term frequency normalized: 1/sqrt(N) where N is the number of terms. We will create a user profile vector of TF-IDF values and multipy this by each recipe preference.

The main idea:

We convert a user's recipe preference into an ingredient preference matrix
We compare each recipe's ingredients to the user's ingredient preferences
We calculate a score for each recipe based on this