Recipes: What makes a recipe ‘healthy’? 🍽️🧂

Authors: Megan Bowen & Jade VanHaitsma

Contact: jadev@umich.edu & megbowen@umich.edu

Introduction

Online recipe websites have now replaced recipe books as the mainstream (and free!) way to find cooking inspiration for one’s next meal. The data set we will be working with is a compilation of recipes and reviews from a popular website, Food.com. This website provides recipes for a variety of cuisines, dating back 15 years (2008), and stopping in 2018. The data sets had 10 years worth of recipe ideas where we would like to investigate what authors should be considered ‘healthy’ in their recipe.

The central question we would like to investigate is, What recipes tend to be healthier? We would like to use data analysis techniques to classify whether or not a recipe would be tagged as ‘healthy’ based on the information in the data. When using a recipe website, whether someone is looking for a healthy lunch, or is on a diet, healthy, delicious recipes are often a main goal for people searching. We want to investigate if the recipes on Food.com are representative and consistent in their healthiness based on their nutrional facts and ingredients in the recipes that creator have deemed ‘healthy’. It is important to know that when one is searching for a healthy recipe, that they are going to be given accurate results.

Introduction to the dataset

The original data set consisted of two data sets which we have merged by the recipe id column. This data set has 234,429 rows and 17 columns. We have chosen the following 16 columns to be included in our DataFrame for analysis. 10 of such columns were not in the original data set, but a subset of the original data cleaned and created from extracting data.

Column Name and Descriptions:

Below are the first 5 rows of our cleaned DataFrame:

name submitted minutes n_steps ingredients n_ingredients n_rating avg_rating calories total fat sugar sodium protein saturated fat carbohydrates year healthy_tag
impossible macaroni and cheese pie 2008-01-01 50 11 [‘cheddar cheese’, ‘macaroni’, ‘milk’, ‘eggs’, ‘bisquick’, ‘salt’, ‘red pepper sauce’] 7 nan 3 386.1 34 7 24 41 62 8 2008 0
impossible rhubarb pie 2008-01-01 55 6 [‘rhubarb’, ‘eggs’, ‘bisquick’, ‘butter’, ‘salt’, ‘sugar’, ‘vanilla’, ‘milk’] 8 nan 3 377.1 18 208 13 13 30 20 2008 1
impossible seafood pie 2008-01-01 45 7 [‘frozen crabmeat’, ‘sharp cheddar cheese’, ‘cream cheese’, ‘onion’, ‘milk’, ‘bisquick’, ‘eggs’, ‘salt’, ‘nutmeg’] 9 nan 3 326.6 30 12 27 37 51 5 2008 0
paula deen s caramel apple cheesecake 2008-01-01 45 11 [‘apple pie filling’, ‘graham cracker crust’, ‘cream cheese’, ‘sugar’, ‘vanilla’, ‘eggs’, ‘caramel topping’, ‘pecan halves’, ‘pecans’] 9 nan 5 577.7 53 149 19 14 67 21 2008 0
midori poached pears 2008-01-01 25 8 [‘midori melon liqueur’, ‘water’, ‘caster sugar’, ‘cinnamon stick’, ‘vanilla pod’, ‘lemon rind’, ‘orange rind’, ‘pear’, ‘mint’] 9 nan 5 386.9 0 347 0 1 0 33 2008 0

Data Cleaning and Exploratory Data Analysis

Data Cleaning

To start the cleaning process, we merged the two data sets, ‘interactions’ and ‘recipes’ on the id column which links the reviews and ratings of recipes with the information about the recipe from the recipe dataset. This allowed us to see the popularity and reviews of the data set.

To follow, we have taken all recipes with rating ‘0’, denoting that there was no rating for that particular interaction, and replaced them with nAn values. Below are the distribution of recipes before and after the change.

After this, we chose to create the column ‘avg_rating` which creates a column for the mean rating of each recipe of all the ratings the recipe has received. This allows us to do calculations on the recipes using one metric as the rating.

Then, we have decided to take the relevant columns from the dataset, dropping the following columns: contributor_id, steps, description, user_id, and recipe_id. We chose to save the irrelevant columns and kept the columns, name,id,minutes,submitted,tags,nutrition,n_steps,n_ingredients,user_id,rating, and avg_rating. These columns will aid in the data analysis.

After selecting our relevant columns, we conducting further cleaning. We knew we were interested in the nutrition column for our health metrics, so we expanded that column into multiple individual columns for values calories, total fats, etc. for the information in nutrition.

We also wanted to look at yearly recipe trends, so we made a year column by extracting the year from the submitted column, which contains the date a recipe was submitted.

Exploratory Data Analysis

Univariate Analysis

For our univariate analysis, we explored the distributions of nutritional information that is often related to the health of a recipe. Here we have histograms describing the distribution of total fats and sugars for all recipes. For many of our variables there are outliers, skewing the dating to be a wide range. We have chosen to crop these graphs in order to get a closer image of the distribution of the graph. We wanted the distribution of total fats, capping the x-axis at 200 (PDV) because this would mean the recipe has 200% more total than the typical daily intake. Values greater than this seemed not to make sense for the variable.

We also wanted to look at the distribution of sugars. For the same reason as total fats, we have attached the full distribution and a cut-off range of [0,200] for the sugar distribution.

Both plots are heavily right-skewed. We also see the medians of these distributions, as indicated by the red dotted line, and middle of the boxplot. The median for saturated fats is 21.0 PDV, and for sugars it is 23.0. This indicates that many of the listed recipes fall within typical daily value ranges for sugar and saturated fat composition.

Bivariate Analysis

This is a plot that describes the distributions of Calories in recipes that are tagged “Unhealthy” (red) vs “Healthy” (green). From this plot, we see the trend that the distribution of unhealthy recipes have a higher number of calories, while the healthy recipes tend to have fewer calories.

Note that due to a high number of outliers, we have cropped the range of the plot to be between 0 and 2000 calories, as that is typical daily intake. Below is the original plot.

We were also interested in the scatter plot between calories and carbohydrates. This gives us insight into how the two nutrients are related, and if there is any pattern when looking at the dots between healthy and unhealthy recipes based on these variables. Looking at the plot, we noticed that there is a slight upward trend between the two variables, and many of the recipes with lower calories are seen as healthy, over the recipes that have high calories.

Interesting Aggregates

Here we’ve aggregated recipes based on if they’ve been tagged ‘healthy’ or not, and are looking at the median nutritional information for each tag category. We can see that recipes per year that were tagged ‘healthy’ have tended to proportionally decrease. In 2019, ther were a sudden drop in recipes tagged ‘healthy’ which we found very interesting as we started our analysis.

year healthy_tag
2008 0.190893
2009 0.188007
2010 0.19753
2011 0.212201
2012 0.18315
2013 0.183017
2014 0.180172
2015 0.124183
2016 0.132353
2017 0.128472
2018 0.121693

Framing a Prediction Problem

Our prediction problem is a binary classification problem. We are hoping to classify recipes as either ‘healthy’ or ‘not healthy’ based on their nutritional content and ingredients – the response variable is healthy_tag. We chose this response variable because given the detailed information on the nutrients and ingredients of these recipes, we felt we could make insightful predictions about general recipe health based on that. When putting together a recipe, the user will have full information about the ingredients and thus nutritional content of the recipe, so it is fair game using predictors from these categories as a basis for whether it would be considered a healthy recipe.

During the time of prediction, we must have the list of ingredients and nutritional information, specifially, the PDV of sodium, carbohydrates, total fats, saturated fats, sugar, and protein, and number of calories. We must know these in order to make a recipe because the contents that make up the meal wil determine their ‘healthiness.’

We have decided to use accuracy as our performance metric because we consider true positives and true negatives equally for our predictions, and do not require the more in-depth analysis offered by F1. While we recognize that F1 often provides a more comprehensive performance metric, especially for imbalanced data, we ultimately decided accuracy was the metric we were able to improve our model upon the most and proceeded with that.

Baseline Model

For our baseline model, we used Logistic Regression. Our features include calories and protein, both of which are quantitative.

Our baseline model’s accuracy is: 0.8080

We do not believe our current model is the best it can be, considering there is room to improve on our performance metric, and we are only using two features (both of which are from nutritional content) as our predictors. Because of this, we also do not think our model would be super generalizable to unseen data.

Our aim is to incorporate more features that could be indicative of the health of a recipe, such as specific ingredients, as well as other nutritional information so that our model can make more informed classifications.

Final Model

In our final model, we included the following nutritional features: calories, total fat, sugar, sodium, protein, saturated fat, carbohydrates. We wanted to use all nutritional information available to inform our model, seeing as just calories and protein do not give the full picture.

Lastly, we engineered 4 one-hot-encoded categorical features derived from ingredients into this model:

After exploring different versions of logistic regression models and decision trees, we ultimately used a Decision Tree for our final model as it performed better than our regression models.

We tuned the hyperparameters using GridSearchCV and got the following optimal values:

Our final model yielded a test accuracy of: 0.8163, which is a slight improvement from our baseline model. We believe incorporating additional nutritional content information as well as factoring in specific ingredient characteristics boosted out model performance by providing more specific insight to each recipe.

We have included visualization of our confusion matrix describing this model’s performance:

It should be noted that our model’s key flaw is that due to the skewness of the classes, we predict many False Negatives. That is, we predict recipes to be ‘unhealthy’ when they are in fact tagged ‘healthy’. This is also a likely cause for why our accuracy did not seem to improve by much even after running hyper parameters and adding relevant features.

One challenge we’ve come to realize is that a recipe being tagged as ‘healthy’ is completely up to the user who submits the recipe. A recipe could be nutritionally healthy, but if a user doesn’t ultimately tag it as ‘healthy,’ our model will fall short. The same goes for recipes that may be nutritionally unhealthy where the user tagged it as ‘healthy.’ Should we continue this analysis further, we would definitely want to refine our model to account for this subjectivity.

December 7, 2024