Protein Prediction Model🍗

By: Kevin Zhang

Introduction

General Introduction

As a frequent-ish gym goer I always wondered how much protein was in my food. It’s very tedious and hard to do without a scale and sometimes really inconvenient making me leave gains on the table. In this project, the primary goal is to predict how much protein is in a recipe to really help maximize my gym gains.

To do this I will:

Merge and Clean the Datasets
Conduct Exploratory Analysis
Frame a Prediction Problem
Create a Base model
Improve Upon the Base Model for my Final Model

Introduction of Dataset

The Data Set contains recipes and ratings from food.com. It was originally scraped and used by the authors of Generating Personalized Recipes from Historical User Preferences (UCSD). The dataset is very large so I will be using data from 2008 and beyond.

There are two datasets, RAW_recipes.csv which contain 86782 rows of recipes and 12 columns with different information such as ‘nutrition’ , while RAW_interactions.csv contains 731927 rows, as each recipe could have more than one review with five columns. However, for my analysis I will focus only on certain columns neccesary for my prediction model.

RAW_recipes.csv:

Column	Description
`name`	Recipe Name
`id`	Recipe ID
`minutes`	Minutes to prepare recipe
`contributor_id`	User ID who submitted this recipe
`submitted`	Date recipe was submitted
`tags`	Food.com tags for recipe
`nutrition`	Nutrition information in the form [calories (#), total fat (PDV), sugar (PDV), sodium (PDV), protein (PDV), saturated fat (PDV), carbohydrates (PDV)]; PDV stands for “percentage of daily value”
`n_steps`	Number of steps in recipe
`steps`	Text for recipe steps, in order
`description`	User-provided description

RAW_interactions.csv:

Column	Description
`user_id`	User ID
`recipe_id`	Recipe ID
`date`	Date of Interaction
`rating`	Rating Given
`review`	Review Text

Data Cleaning and Exploratory Data Analysis

Cleaning

The first step in our data cleaning process is converting rating of 0 to NaN. This is because when commenters review recipes and do not give a rating, the site by default, assigns the rating to be 0 when a normal should fall within the 1 - 5 range. Thus we ignore these 0 ratings.

First, I changed the column recipe_id to id, so i could conduct a left merge onto the id column. Now one dataframe, I dropped any duplicate columns. Second, I found the average rating per recipe and created a new column avg_rating. Third, I extracted nutrition information from the nutrition column creating new columns in calories , protein , and total_fat . Finally, I dropped no longer neccessary columns such as user_id , date, etc.

Final DataFrame:

minutes	tags	nutrition	n_steps	ingredients	rating	avg_rating	calories	protein	total_fat	protein_calorie_ratio
40	[‘60-minutes-or-less’, ‘time-to-make’, ‘course’, ‘main-ingredient’, ‘preparation’, ‘for-large-groups’, ‘desserts’, ‘lunch’, ‘snacks’, ‘cookies-and-brownies’, ‘chocolate’, ‘bar-cookies’, ‘brownies’, ‘number-of-servings’]	[138.4, 10.0, 50.0, 3.0, 3.0, 19.0, 6.0]	10	[‘bittersweet chocolate’, ‘unsalted butter’, ‘eggs’, ‘granulated sugar’, ‘unsweetened cocoa powder’, ‘vanilla extract’, ‘brewed espresso’, ‘kosher salt’, ‘all-purpose flour’]	4	4	138.4	3	10	0.0216763

Univariate Analysis

The following graph shows the distribution of the amount of protein content among all recipes with 0 - 200 grams. The graph is skewed left as many recipes, such as desserts, have little to no protein. Making the median amount of protein being 18 grams. Understanding the distribution of protein gives valuable insight into protein trends.

This graph shows the top eight most common tags. Understanding the distribution of the top 8 tags allows us to have a good understanding for when we see the relationship between tags and protein content later.

Bivariate Analysis

Building off the last graph, we can see the average amount of protein per tag. We can see that certain tags such as ‘Main Ingredient’ and ‘cuisine’ have the highest, while ‘easy’ has the lowest. This makes sense as many main dishes feature some type of protein, while on the other hand, easy dishes usually don’t incorporate as much protein due to easy usually meaning shorter cook time and proteins usually take longer to cook and prep.

Lastly, we utilized a scatter plot to see the correlation between protein and number of steps. And see that there is a slight negative correlation between number of steps and protein content. Again, this makes sense as many baked goods have lots of steps but wouldn’t have much protein at all.

Aggregations

One interesting relationship I wanted to look at is the relationship between fats and protein. As many foods with protein come with many fats. We can see below with the mean and median there is a heavy correlation between lots of fats and very high protein content. This will be very useful to think about as we build our model.

fat_category	mean	median	count
Very High	91.6309	75	19593
High	54.0364	50	36956
Medium	35.185	28	57640
Low	18.5354	10	103535

Imputations

I did not need to do any missing value imputation as all of my data that I would later use for my model had zero nan values

Framing a Prediction Problem

Using all of my previouse analysis I will attempt to predict protein content using Tags, total fats, and n_steps to make. I will do this using a ridge regression model. I will be using the MSE to measure my accuracy in this process

Baseline Model

The baseline mode will be a ridge regression, using standard scalar for quantitative features total_fats and n_steps. I chose ridge regression not only because it reduces the variance of the model, making it less sensitive to extreme values but also because it helps deal with possible multi collinearity in the tags column when we OneHotEncode it as there may be overlap in meaning. Additionally, I OneHotEncoded Tags, because it is a categorical feature, so the model could use it to make predictions. Lastly, I used sklearn’s train_test_split of 20% test with 80% train split. This way we can see how our training data handles unseen data, along with detecting possible overfitting of our model.

Feature	Variable Type
`tags`	Nominal - OneHotEncoded
`total_fats`	Quantitative - Standard Scaler
`n_steps`	Quantitative - Standard Scaler

Test Set	MSE
Train	1596.29
Test	1688.77

We used mean square error to measure our accuracy and with our ridge model we had a training mse of 1596.29 and a test mse of 1688.77. This is already a solid model, not only because we are working with such a big data set, but because we also have heavy skewing in our dataset with the max protein in a recipe to be over 4000 while our median is just 18. But lets try to make an even better model!

Final Model

For my final model, two new features I incorporated were: the log_total_fat and steps_per_fat. I chose to take the log of total fat because total fat was heavily skewed right with the maximum being over 3000 grams while most recipes have only 20grams or less. So by taking the log we can reduce the skew and create a more nominal distribution which will allow us to make better predictions. For steps_per_fat I did n_steps/(total_fat + 1). I thought it would help the model capture a more nuanced relationship between recipe complexity and nutritional density. For example, a burger would have very few steps but have a above average fat would have low steps_per_fat. On the other hand, a salad with ten steps and a low fat content would have high steps_per_fat. Using this way of thinking, high-effort, low-fat recipes might be protein-rich (like lean meat stir-fries or vegan stews) while low-effort, high-fat might be fatty but low in protein (like buttery toast).

I used GridSearchCV to tune the regularization strength parameter alpha over the values [0.1, 1.0, 10.0, 100.0], using 5-fold cross-validation on the training set and neg_mean_squared_error as the scoring metric. The best-performing hyperparameter is 10, showing that the model is not overfitted but also not blind in picking up trends.

My final result showed a slight lower test MSE and RMSE than the baseline, showing that the newly engineered features (log_total_fat and1 steps_per_fat ratio) subtly contributed some predictive signal, but that the baseline model was already performing well.

Model	Train MSE	Test MSE
Base	1596.29	1688.77
Final	1558.91	1642.71