A Food Data Analysis: Recipe Ratings & Calories

Authors: Clayton Hammen Tan, Varick Janiro Hasim

Overview

The goal of this project is to analyze which recipe factors impact calories and to predict a recipe’s calories accurately based on its other attributes (e.g., average rating, number of steps, number of ingredients). Using two datasets (recipes and user interactions), we clean and explore the data, test hypotheses, build predictive models, and ensure fairness across different recipe categories.

Introduction

This project investigates how various factors contribute to a recipe’s calories using data from Food.com. The datasets include:

By merging these datasets, we explore trends in calorie content and predict calories based on recipe attributes.

Research Question

What factors influence the calories and average rating of a recipe?

This question helps determine if more complex recipes tend to have higher calories and different satisfaction levels. The insights can guide home cooks, food bloggers, and recipe developers.

Dataset Description

Recipes Dataset

Column Name Description
name Name of the recipe, providing a unique identifier for each dish.
id Unique Recipe ID.
minutes Total preparation time (in minutes).
contributor_id ID of the user who submitted the recipe.
submitted Date when the recipe was submitted.
tags Categories associated with the recipe (e.g., cuisine or meal type).
nutrition Nutritional information including calories, fat, protein, and carbohydrates.
n_steps Number of steps in the recipe instructions.
steps Detailed recipe instructions.
description User-provided description of the recipe.
ingredients List of ingredients used.
n_ingredients Total number of ingredients used.
average_rating Average rating, calculated from user interactions.

Interactions Dataset

Column Name Description
user_id ID of the user providing the rating or review.
recipe_id ID of the reviewed recipe.
date Date of the interaction (review or rating).
rating Rating given by the user.
review Textual review provided by the user (optional).

Importance of the Study

Understanding which recipe attributes influence calories has practical and academic benefits. It helps recipe developers and everyday cooks estimate recipe calories and demonstrates robust data analysis and predictive modeling techniques applicable to other recommendation systems.

Data Cleaning and Exploratory Data Analysis

Data Cleaning

The following cleaning steps were performed:

  1. Merged the Datasets:
    The recipes and interactions datasets were merged using a left join on id and recipe_id so that all recipes were retained.

  2. Replaced Invalid Ratings:
    Ratings of 0 in the interactions dataset were replaced with NaN, treating them as missing rather than true ratings.

  3. Calculated Average Ratings:
    Average ratings were computed for each recipe (grouped by name) and added as a new column average_rating, rounded to two decimals.

  4. Extracted Nutritional Information:
    The nutrition column (a string) was split into separate columns for calories, total_fat, sugar, sodium, protein, saturated_fat, and carbohydrates.

  5. Created Additional Binary Columns:
    • simple: True if the number of ingredients ≤ 9.
    • easy: True if the number of steps ≤ 9.
    • healthy: True if the ratio of saturated fat to calories ≤ 0.10.
  6. Removed Outliers:
    Outliers (e.g., recipes with 40,000 calories) were removed.

Effect on Analysis:

Univariate Analysis

Plot 1: Distribution of Number of Ingredients

This histogram shows the typical number of ingredients per recipe.

Plot 2: Distribution of Calories in Recipes

This density-based histogram shows the calorie content across recipes.

Bivariate Analysis

Scatter Plot: Calories vs. Average Rating

This plot suggests that while recipes with high ratings span a range of calorie values, calorie count alone may not drive ratings.

Box Plot: Average Rating by Number of Steps

This box plot indicates that recipes with fewer preparation steps have more consistent average ratings, and those with more have more varying average ratings.

Interesting Aggregates

Pivot Table: Calories and Ratings by Cooking Time (Minutes)

The table below shows how cooking time correlates with satisfaction and caloric content:

average_rating calories
5 653.1
4.73875 195.705
4.72001 223.325
4.6944 253.525
4.53713 271.961

Assessment of Missingness

NMAR Analysis: Understanding Missingness in average_rating

The average_rating field exhibits NMAR (Not Missing At Random) behavior, meaning its missingness is driven by factors not explicitly captured in the dataset.

Why is average_rating NMAR?

How Could We Make It MAR?

Additional data (e.g., user interaction frequency, contextual details, engagement metrics) could link the missingness to observable factors, allowing it to be treated as MAR (Missing At Random).

Missingness Dependency: Investigating Relationships

Permutation tests were conducted to determine if the missingness of average_rating depends on other recipe features. For each test:

Detailed Results

  1. Missingness vs. calories
    • Observed Test Statistic: 19.308042
    • P-value: 0.000

    Conclusion: Depends on calories.

  2. Missingness vs. healthy
    • Observed Test Statistic: 0.026784
    • P-value: 0.004

    Conclusion: Depends on healthy.

  3. Missingness vs. simple
    • Observed Test Statistic: 0.024634
    • P-value: 0.026

    Conclusion: Depends on simple.

  4. Missingness vs. easy
    • Observed Test Statistic: 0.073634
    • P-value: 0.000

    Conclusion: Depends on easy.

  5. Missingness vs. sugar
    • Observed Test Statistic: 9.587501
    • P-value: 0.000

    Conclusion: Depends on sugar.

  6. Missingness vs. protein
    • Observed Test Statistic: 0.993006
    • P-value: 0.120

    Conclusion: Does not depend on protein.

  7. Missingness vs. sodium
    • Observed Test Statistic: 0.828399
    • P-value: 0.622

    Conclusion: Does not depend on sodium.

  8. Missingness vs. n_ingredients
    • Observed Test Statistic: 0.259784
    • P-value: 0.000

    Conclusion: Depends on n_ingredients.

  9. Missingness vs. n_steps
    • Observed Test Statistic: 1.471825
    • P-value: 0.000

    Conclusion: Depends on n_steps.

  10. Missingness vs. saturated_fat
    • Observed Test Statistic: 3.117959
    • P-value: 0.000

    Conclusion: Depends on saturated_fat.

Summary Table

Tested Column Observed Statistic P-Value Conclusion
calories 19.308042 0.000 Depends on calories
healthy 0.026784 0.004 Depends on healthy
simple 0.024634 0.026 Depends on simple
easy 0.073634 0.000 Depends on easy
sugar 9.587501 0.000 Depends on sugar
protein 0.993006 0.120 Does not depend on protein
sodium 0.828399 0.622 Does not depend on sodium
n_ingredients 0.259784 0.000 Depends on n_ingredients
n_steps 1.471825 0.000 Depends on n_steps
saturated_fat 3.117959 0.000 Depends on saturated_fat

Summary:
Missingness in average_rating is significantly associated with calories, healthy, simple, easy, sugar, n_ingredients, n_steps, and saturated_fat (p < 0.05) but not with protein or sodium.

Hypothesis Testing

Objective

Determine whether a recipe’s number of preparation steps affects its average rating.

Hypotheses

Test Statistic & Significance Level

The test statistic is the absolute difference in the mean average rating between:

Significance Level: 0.05

Methodology

  1. Calculate the observed test statistic (absolute difference in means).
  2. Shuffle the group labels (≤9 steps vs. >9 steps) 1,000 times to simulate the null hypothesis.
  3. Compute the p-value as the proportion of permuted statistics ≥ the observed statistic.

Results & Visualization

Conclusion

Since the p-value (~0.87) is greater than 0.05, we fail to reject the null hypothesis. This suggests that the number of preparation steps does not significantly influence the average rating.

Framing a Prediction Problem

Prediction Problem Statement

Predict the total calories (calories) of a recipe based on its characteristics:

Predicting calories helps understand how complexity and nutritional content contribute to a recipe’s caloric value, guiding recipe creation and dietary choices.

Type of Prediction Problem

This is a regression problem since the target variable (calories) is continuous (ranging from 58 to 1036).

Response Variable

Calories – Represents the nutritional content and indirectly the healthiness of the recipe.

Features Used for Prediction

All features are known at the time of prediction.

Evaluation Metric

We use:

Visual Exploration of the Problem

  1. Distribution of Calories:

    (Right-skewed distribution with most recipes having lower calories.)

  2. Scatter Plot: n_ingredients vs. Calories:

    (No clear relationship observed.)

  3. Scatter Plot: n_steps vs. Calories:

    (Most recipes have ≤40 steps; relationship is unclear.)

  4. Scatter Plot: Protein vs. Calories:

    (Positive relationship observed.)

  5. Scatter Plot: Sodium vs. Calories:

    (No clear relationship.)

  6. Scatter Plot: Saturated Fat vs. Calories:

    (Positive relationship observed.)

  7. Scatter Plot: Sugar vs. Calories:

    (Positive relationship observed.)

Importance

Predicting calories benefits:

Baseline Model

The baseline model is a Linear Regression model built using a preprocessing pipeline.

Model Features

Numerical Features

(Standardization was not applied since it does not affect predictions in linear regression.)

Categorical Feature

Model Performance

Metrics:

Coefficients:

The results suggest that n_ingredients, protein, and saturated_fat are significant predictors of calorie content.

Model’s Strengths and Weaknesses

Strengths:

  1. Reproducible and scalable.
  2. Effective handling of categorical data via one-hot encoding.
  3. Interpretability of coefficients.

Weaknesses:

  1. Prone to outliers.
  2. Assumes linear relationships, which may not hold for all predictors.

Final Model

Added Features

  1. complexity: n_steps + n_ingredients
    (Captures overall recipe complexity.)
  2. bad_nutrition: sugar + saturated_fat + sodium
    (Captures cumulative negative nutritional impact.)

Modeling Algorithm

A Random Forest Regressor was used because it:

Hyperparameter Tuning

Using GridSearchCV (3-fold CV), the best hyperparameters were:

Performance Comparison

Baseline Model:

Final Model:

Conclusion

The final model’s lower MAE and higher R² demonstrate improved accuracy. The enhancements are attributed to the new features and the non-linear modeling capability of the Random Forest Regressor.

Fairness Analysis

A fairness analysis was conducted by comparing model performance (MAE) between two groups:

Hypotheses

Test Statistic

The absolute difference in MAE between the two groups.

Significance Level

α = 0.05

Results

Since the p-value is below 0.05, we reject H₀, indicating that the model’s performance differs between Simple and Complex Recipes. This suggests potential unfairness that warrants further investigation.

As the permutation test’s p-value of 0 is less than the chosen significance level of 0.05, we reject the null hypothesis. Hence, the observed difference in MAE is statistically significant, and it indicates that the model’s performance does differ between the two groups recipes, performing slightly better for one group of recipes. Hence, there is potential unfairness in the model between Simple and Complex Recipes, suggesting areas for improvement.

The histogram below illustrates the distribution of permutation test statistics under the null hypothesis, with the observed difference marked by the red dashed line.