Recipe Nutrition Analysis

By: Ilan Breines
Contact: ilanbrei@umich.edu

Introduction

This analysis explores recipe data from Food.com, one of the largest recipe-sharing platforms on the internet. The dataset contains recipes and their nutritional information, sourced from user submissions since 2008. Our central question is: Can we predict a recipe’s calorie content based on its other nutritional information and characteristics?

Why This Matters

Understanding the relationship between different nutritional components and calorie content is crucial for:

• Home cooks trying to plan healthy meals
• People managing their dietary intake
• Recipe developers working on health-conscious alternatives
• Anyone interested in understanding how different nutrients contribute to overall caloric content

Dataset Overview

The dataset contains 215,772 rows with detailed nutritional and preparation information as well as recipe ratings.

Relevant columns for our analysis:

• calories: Number of calories per serving (our prediction target)
• total_fat: Total fat content as percentage of daily value (PDV)
• protein: Protein content as PDV
• carbohydrates: Carbohydrate content as PDV
• sugar: Sugar content as PDV
• sodium: Sodium content as PDV
• saturated_fat: Saturated fat content as PDV
• n_steps: Number of steps in the recipe preparation
• n_ingredients: Number of ingredients used

Data Cleaning and Exploratory Data Analysis

Data Cleaning Process

Several critical cleaning steps were necessary to prepare yhe recipe data for analysis:

1. Merging Recipe and Rating Data
   • Performed a left join between recipes and ratings datasets to preserve all recipes even without ratings
   • This ensures all recipes are included in our analysis, regardless of review status
2. Handling Cooking Times
   • Filtered out recipes with preparation times beyond the 1st and 99th percentiles
   • This removes likely data entry errors (e.g., recipes claiming to take several days) and extremely unusual recipes
   • Keeps our analysis focused on typical home cooking scenarios
3. Rating Processing
   • Replaced ratings of 0 with NaN values since 0 ratings on Food.com indicate missing ratings rather than actual zero-star reviews
   • Calculated average rating per recipe using transform(‘mean’) to maintain data alignment
   • Counted number of ratings per recipe to understand review frequency
   • Filled missing average ratings with 0 to indicate unrated recipes
4. Deduplication
   • Used drop_duplicates(subset=’id’) to keep only one instance of each unique recipe
   • This prevents popular recipes with many reviews from being overrepresented in our analysis
   • Ensures our calorie predictions are based on distinct recipes
5. Nutritional Information Extraction
   • Split the combined ‘nutrition’ string into separate columns for each component
   • Converted nutrition values from strings to numeric format for analysis
   • Original format was a list-like string: ‘[calories, total_fat, sugar, sodium, protein, saturated_fat, carbohydrates]’
6. Tag Processing
   • Cleaned recipe tags by removing brackets, quotes, and extra whitespace
   • Split tag strings into lists for potential categorical analysis
7. Outlier Removal
   • Removed recipes with nutritional values above the 99th percentile
   • This step ensures our analysis isn’t skewed by potential data entry errors
   • Applied to: calories, total fat, sugar, sodium, protein, saturated fat, and carbohydrates

Missing Values Analysis

I performed a programmatic check for missing values in all relevant columns and found no missing values. Therefore, no imputation was necessary for our analysis. This completeness is likely due to Food.com’s submission requirements.

Here’s what the cleaned dataset looks like (displaying only columns relevant to our analysis):

name	id	minutes	contributor_id	submitted	tags	nutrition	n_steps	steps	description	ingredients	n_ingredients	user_id	recipe_id	date	rating	review	avg_rating	n_ratings	calories	total_fat	sugar	sodium	protein	saturated_fat	carbohydrates
1 brownies in the world best ever	333281	40	985201	2008-10-27	[‘60-minutes-or-less’, ‘time-to-make’, ‘course’, ‘	[138.4, 10.0, 50.0, 3.0, 3.0, 19.0, 6.0]	10	[‘heat the oven to 350f and arrange the rack in th	these are the most; chocolatey, moist, rich, dense	[‘bittersweet chocolate’, ‘unsalted butter’, ‘eggs	9	386585	333281	2008-11-19	4	These were pretty good, but took forever to bake.	4	1	138.4	10	50	3	3	19	6
1 in canada chocolate chip cookies	453467	45	1848091	2011-04-11	[‘60-minutes-or-less’, ‘time-to-make’, ‘cuisine’,	[595.1, 46.0, 211.0, 22.0, 13.0, 51.0, 26.0]	12	[‘pre-heat oven the 350 degrees f’, ‘in a mixing b	this is the recipe that we use at my school cafete	[‘white sugar’, ‘brown sugar’, ‘salt’, ‘margarine’	11	424680	453467	2012-01-26	5	Originally I was gonna cut the recipe in half (jus	5	1	595.1	46	211	22	13	51	26
412 broccoli casserole	306168	40	50969	2008-05-30	[‘60-minutes-or-less’, ‘time-to-make’, ‘course’, ‘	[194.8, 20.0, 6.0, 32.0, 22.0, 36.0, 3.0]	6	[‘preheat oven to 350 degrees’, ‘spray a 2 quart b	since there are already 411 recipes for broccoli c	[‘frozen broccoli cuts’, ‘cream of chicken soup’,	9	29782	306168	2008-12-31	5	This was one of the best broccoli casseroles that	5	4	194.8	20	6	32	22	36	3
millionaire pound cake	286009	120	461724	2008-02-12	[‘time-to-make’, ‘course’, ‘cuisine’, ‘preparation	[878.3, 63.0, 326.0, 13.0, 20.0, 123.0, 39.0]	7	[‘freheat the oven to 300 degrees’, ‘grease a 10-i	why a millionaire pound cake? because it’s super	[‘butter’, ‘sugar’, ‘eggs’, ‘all-purpose flour’, ‘	7	813055	286009	2008-04-09	5	don’t let the calories and fat grams scare you off	5	1	878.3	63	326	13	20	123	39
2000 meatloaf	475785	90	2202916	2012-03-06	[‘time-to-make’, ‘course’, ‘main-ingredient’, ‘pre	[267.0, 30.0, 12.0, 12.0, 29.0, 48.0, 2.0]	17	[‘pan fry bacon , and set aside on a paper towel t	ready, set, cook! special edition contest entry: a	[‘meatloaf mixture’, ‘unsmoked bacon’, ‘goat chees	13	2.20436e+06	475785	2012-03-07	5	Delicious!!!!! – the goat cheese made the differe	5	2	267	30	12	12	29	48	2

Calorie Distribution and Key Relationships

The distribution of calories across recipes shows a right-skewed pattern, with most recipes containing moderate calorie counts. This aligns with what we might expect from real-world recipes, where extreme calorie counts are less common.

Our analysis revealed strong correlations between calories and other nutritional components:

The scatter plot demonstrates a strong positive correlation between total fat content and calories, suggesting that fat content is a strong predictor of a recipe’s calorie count.

Here’s how nutritional values and recipe characteristics vary across calorie quintiles:

calories	total_fat	protein	carbohydrates	sugar	sodium	saturated_fat	n_ingredients	n_steps
0-20%	4.89	4.75	3.1	20.98	7.49	5.74	7.42	7.96
20-40%	11.9	14.32	6.76	35.88	12.66	14.05	8.66	9.18
40-60%	19.79	25.46	9.54	44.43	18.7	24.07	9.32	9.87
60-80%	30.41	37.88	12.8	54.82	25.5	38.49	9.91	10.79
80-100%	57.31	58.81	18.79	68.64	38.04	69.57	10.71	12.08

This table reveals several interesting patterns:

• As calories increase, all nutritional components tend to increase proportionally
• Higher-calorie recipes tend to be more complex, with more ingredients and steps
• The jump in nutritional values is particularly dramatic in the highest calorie quintile (80-100%)
• Sugar content shows the smallest relative increase across quintiles compared to other nutrients

Framing a Prediction Problem

I formulated my analysis as a regression problem to predict recipe calorie content from other nutritional and recipe characteristics.

Problem Statement

• Task Type: Regression (predicting a continuous numerical value)
• Response Variable: Calories per serving
• Features Available at Prediction Time:
• Other nutritional information (fats, proteins, carbohydrates, etc.)
• Recipe characteristics (number of steps, number of ingredients)
• Note: We can’t use ratings or reviews as features since these would only be available after a recipe is published

Why Predict Calories?

I chose calories as the prediction target because:

1. It’s a critical nutritional metric that directly impacts meal planning and dietary decisions
2. Accurate calorie information is essential for recipe websites and health-focused applications
3. Unlike subjective measures like ratings, calories can be objectively measured and validated

Evaluation Metric Choice

I selected Mean Absolute Error (MAE) as the evaluation metric because:

1. Interpretability: MAE shows the average calorie difference between predictions and actual values. “Our predictions are off by 77 calories” is more meaningful than a squared error.

2. Robustness: MAE handles outliers (like very light snacks vs. heavy meals) better than squared error metrics.

3. Linear Scale: A 100-calorie error isn’t necessarily 4 times worse than a 50-calorie error. MAE’s linear scale reflects this better than MSE’s quadratic scale.

Alternative metrics considered:

• Mean Squared Error (MSE): Overpenalizes large errors
• Root Mean Squared Error (RMSE): Same issue as MSE despite using calorie units
• Mean Percentage Error (MPE): Misleading for low-calorie recipes where small errors appear large

  Baseline Model

My baseline model aims to predict recipe calories using just two fundamental nutritional features:

Features

I selected two quantitative features for my initial model:

1. Total Fat (PDV)
2. Protein (PDV)

Both features are continuous numerical values representing the Percentage of Daily Value (PDV) of these nutrients. No encoding was necessary as these features were already in numerical format after our initial data cleaning.

Feature Breakdown:

• Quantitative features: 2 (total fat, protein)

Model Performance

Using a simple linear regression, our baseline model achieved:

• Mean Absolute Error (MAE): 77.35 calories
• For context:
• Average recipe in dataset: 332.59 calories
• Error represents about 23.3% of mean calorie content

Model Performance

Strengths:

1. Relatively simple and interpretable
2. Uses just two readily available nutritional metrics
3. Average error of 77 calories is reasonable for meal planning purposes

Limitations:

1. Ignores other potentially useful nutritional information (carbohydrates, sugar, etc.)
2. Doesn’t account for recipe complexity (number of steps, ingredients)
3. Doesn’t consider interactions between nutritional components

While the baseline model provides a reasonable starting point, there’s clear room for improvement.

Final Model

Feature Engineering Rationale

For my final model, I initially tried adding engineered features such as nutritional ratios, but found the basic nutritional components were most effective. This makes scientific sense because:

• Calories are directly determined by macronutrients
• Basic nutrients already capture the key relationships

Ratios Weren’t Helpful because:

• Engineered ratios like protein-to-fat were redundant since they’re derived directly from existing features
• The linear relationships between macronutrients and calories meant simple features worked best

Additional Selected Features

1. Basic Nutritional Components:
   • Carbohydrates: A primary source of calories alongside fats and proteins
   • Sugar: Indicates caloric density from simple carbs
   • Sodium: Can indicate preservation methods and cooking styles
   • Saturated Fat: Distinguishes between different types of fats with varying caloric impacts
2. Recipe Complexity Metrics:
   • Number of Steps: More preparation steps often indicate more complex cooking methods that can affect caloric content
   • Number of Ingredients: More ingredients typically suggest more complex dishes with potentially higher calorie content

Modeling Choices

I selected Lasso regression as our final model for several reasons:

• Performs automatic feature selection through L1 regularization
• Maintains interpretability of coefficients
• Can handle correlated features like the engineered nutritional ratios

Hyperparameter Selection

For our Lasso regression model, we’ll tune the alpha parameter, which controls the strength of regularization. This is important because:

1. With many nutritional features that may be correlated, we need to find the right balance between fitting the data and preventing overfitting
2. The alpha parameter will help identify which features are most important for prediction by potentially shrinking less important coefficients to zero

Hyperparameter Tuning

• Used GridSearchCV with 5-fold cross-validation
• Tuned Lasso’s alpha parameter across 50 values from 10^-4 to 1
• Best alpha = 1.0000, indicating strong regularization was beneficial
• This suggests some features were less important for prediction

Model Performance

The final model achieved an MAE of 10.77 calories, compared to the baseline’s 77.35 calories, an 86% improvement. I am quite happy with this outcome.

Insights from feature coefficients:

1. Primary Nutritional Predictors:
   • Carbohydrates (11.16), total fat (5.61), and protein (2.18) were the strongest predictors
   • This aligns with the scientific understanding that these macronutrients directly determine caloric content
2. Minimal Impact Features:
   • Sugar and sodium had near-zero coefficients
   • Number of ingredients showed a small negative coefficient
3. Model Characteristics:
   • The moderate Lasso alpha (0.0869) indicates some regularization was beneficial
   • The model effectively focused on the core nutritional components that directly determine calories