Calorie Counter: What Impacts a Recipe’s Calorie Count?

by Jamila Nagarwala and Mehek Gupta jnagar@umich.edu and mehekgup@umich.edu

Introduction

This project analyzes the ‘Recipes and Ratings’ dataset from food.com, which contains information about various recipes along with user ratings and reviews. The dataset used is relevant for people interested in cooking, nutrition, and recipe recommendations. The data comes in two parts:

  1. Recipes Dataset: Contains details about recipes, including preparation time, number of ingredients, steps, and nutritional information.
  2. Interactions Dataset: Includes user reviews and ratings, providing insights into recipe performance.

Question/Why It Matters

The project is centered around the question: “What types of recipes tend to be healthier, or in other words what types of recipes tend to have a lower calorie count?” This is an important question to evaluate because with increasing awareness about health and nutrition, many people are seeking healthier recipe options to support their lifestyle. Understanding what components relate to a recipe being lower in calories can help individuals make better dietary choices. This analysis aims to identify key recipe factors, such as preparation time, number of ingredients, and nutritional components, that affect the total calorie count of a recipe.

Dataset Information

The number of rows in the dataset is 231637, the number of columns before cleaning was 12 and after cleaning is 15. The relevant column names and descriptions to our analysis are as follows:

Recipes Dataset:

Data Cleaning and Exploratory Data Analysis

Data Cleaning

  1. The recipes data set was merged with the interaction data set with the ‘id’ of the recipe. This added the rating columns into the recipe data set.

  2. In the merged dataset, the ratings of ‘0’ were filled with NaN, since the rating scale was from 1 to 5, which does not include the value of 0. The recipes with ratings of ‘0’ had missing rating data.

  3. The average rating per recipe was calculated and added as a new column to the dataframe (as a Series).

  4. Nutritional information was extracted from the ‘nutrition’ column into separate columns for calories, fat, sugar, sodium, protein, etc.

  5. The missing values were handled by not imputing them since there was no accurate and reliable way to estimate them.

  6. Columns unnecessary to our analysis were dropped such as the tag column, ingredient column, minutes column, and more.

print(recipes.head().to_markdown(index=False))
name n_ingredients average_rating calories (#) total_fat (PDV) sugar (PDV) sodium (PDV) protein (PDV) saturated_fat (PDV) carbohydrates (PDV)
arriba baked winter squash mexican style 7 5 51.5 0 13 0 2 0 4
a bit different breakfast pizza 6 4.66667 173.4 18 0 17 22 35 1
all in the kitchen chili 13 4 269.8 22 32 48 39 27 5
alouette potatoes 11 4.5 368.1 17 10 2 14 8 20
amish tomato ketchup for canning 8 5 352.9 1 337 23 3 0 28

Univariate Analysis

Distribution of Calories

This histogram, which shows the distribution of the number of calories across all recipes that are 3000 calories or lower, in order to eliminate extreme outliers, stipulates that the majority of the recipes fall in the 0 to 500 calories range, with a peak at around 150 to 190 calories. This right skewed distribution indicates that while most recipes are moderately low in calories, there still exists a long tail of higher calorie recipes, which helps us understand what is considered a relatively “low-calorie” recipe in terms of this dataset.

Distribution of Protein

This histogram, which shows the distribution of the protein amount by PDV across all recipes that are 150% PDV or lower, in order to eliminate extreme outliers, stipulates that the main peak of the recipes fall in the 0% to 7% PDV range (~60,000 recipes), and the majority of the recipes (95% of them) fall in the 5% to 50% PDV range. This right skewed distribution indicates that most recipes are relatively low to moderate in protein, but there still exists a long tail of higher protein recipes.

Bivariate Analysis

Distribution of Calories by the Number of Ingredients

This box plot, which shows the distribution of the number of ingredients across all recipes that are 20 ingredients or lower by the number of calories across all recipes that are 3000 calories or lower, in order to eliminate extreme outliers, shows a trend: recipes with higher ingredient counts generally have slightly higher calorie content.

Distribution of Saturated Fat by Calories

The scatter plot illustrates the distribution of saturated fat (PDV) by calories, using a filtered recipe dataset that only contains recipes that are 3000 calories or lower, in order to eliminate extreme outliers. The plot suffers from overplotting, since there are so many points on top of one another making it hard to see the general trend.

The box plot is used to address the overplotting issue observed in the scatter plot and displays the distribution of saturated fat (PDV) across different calorie ranges, filtered so that only recipes that are 3000 calories or lower, in order to eliminate extreme outliers, are considered. The graph shows a trend: recipes with higher calorie counts have higher saturated fat values.

Interesting Aggregates

print(pivot_table.to_markdown(index=False))
ingredient_range calories (#) carbohydrates (PDV) protein (PDV) sodium (PDV) sugar (PDV) total_fat (PDV)
0-9 424.793 14.2776 28.7566 27.4052 84.6468 32.1586
10-19 539.225 17.2994 42.5371 33.448 83.6886 41.2413
20-29 720.483 20.7042 65.8122 55.4133 86.0834 57.2432
30-39 1342.26 36.3148 99.7037 147.667 162.593 122.648

This pivot table displays nutritional information for different ranges of ingredient counts. The data is filtered to include only recipes with 40 or fewer ingredients and is grouped into intervals of 10 ingredients. From the pivot table, we observe that recipes with more ingredients generally have higher average calorie counts and nutritional values. This allows us to identify which recipes are typically healthier, based on their lower calorie content and higher nutritional value.

Imputation

Missing rating values were not imputed because the ratings of ‘0’ were invalid due to the rating scale being 1-5, there was no reliable way to estimate missing ratings, and the missing ratings were likely missing at random. So these recipes didn’t skew the analysis, ratings of ‘0’ were imputed with NaN.

Framing a Prediction Problem

The goal of this analysis is to predict calorie count of recipes based on various features from the dataset.

Problem Type

This is a regression problem since calorie count is a continuous numerical variable, the differences between calorie values can be quantitatively measured, and the goal is to predict precise numerical outcomes.

Response Variable/Features

The response variable is calories, which was extracted from the ‘nutrition’ column. This was chosen as the response variable because the number of calories in a recipe can be used to answer the initial question about which recipes tend to be healthier.

Features

Relevance: The selected features directly relate to key nutritional factors influencing calorie content.

Availability: These features are always available at the time of prediction, as they are from a recipe’s nutritional information. The information known at the “time of prediction” include number of ingredients, preparation time, and nutritional components that may correlate with calories.

Evaluation Metrics

The evaluation metric is Mean Squared Error (MSE). This was chosen because it penalizes large errors more heavily, it’s commonly used for regression problems, and the square units match people’s intuitive understanding of calorie differences. We will also use R² Score as it explains the proportion of variance in calorie count that can be predicted by the selected feature.

Other possible metrics for regression, such as Mean Absolute Error (MAE) or Root Mean Squared Error (RMSE), were considered but ultimately not used as primary metrics. MAE treats all errors equally and doesn’t emphasize larger deviations, which makes it less sensitive to outliers. RMSE, while interpretable in the same units as the target variable (calories), carries the same penalty characteristics as MSE and would lead to similar conclusions, so we chose to report the simpler MSE for consistency and ease of comparison.

Baseline Model

Model Description

The baseline model is a Linear Regression model that predicts the calorie content of a recipe based on two quantitative features that are related to calories: protein content (PDV) and number of ingredients.

Features

  1. Quantitative Features:
    • protein (PDV): Protein content in Percent Daily Value (PDV%).
    • number of ingredients: The number of ingredients in the recipe.
    • Protein is a quantiatvie continuous feature, while number of ingredients is a quantitative discrete feature.
    • Both features are numerical, therefore they require no special encoding but required standardization. Both feature were standardized using StandardScaler to make the values comparable, as linear regression models are sensitive to the magnitude of the features.
  2. Response Variable:
    • calories: Number of calories, the response variable, measured as a continuous quantitative value.

Model Performance

  1. Mean Squared Error (MSE):
    • Train MSE: 93680.658
    • Test MSE: 93217.755

    Interpretation: The model’s predictions differ significantly from the actual calorie values, suggesting that this basic model has difficulty accounting for the variability in calorie content.

  2. R² Score (Coefficient of Determination):
    • Train R²: 0.3656
    • Test R²: 0.3749

    Interpretation: The model explains only ~37% of the variance in calorie values.

Is This a Good Baseline Model?

This is a weak baseline model, as it has a low explantory power (low R² score) and prediction accuracy (high MSE). The low R² score suggests that the features chosen (protein and number of ingredients) are insufficient to predict calories effectively on their own and the high MSE suggests that only utilizing protein and number of ingredients leads to high error, meaning it does not predict calories well.

Final Model

Model Description

The final model uses additional engineered features in order to better predict the calorie content of a recipe based on multiple nutritional information and number of ingredients. It utilizes a more advanced regression model and hyper-tunes parameters to increase its predictive performance.

Features

  1. Quantitative Features:
    • number of ingredients: the number of ingredients in the recipe. Based on our EDA analysis, calories in a recipe tends to increase as number of ingredients increases indicating there is some correlation between the number of ingredients and calories.
    • total_fat (PDV): Total fat content in Percent Daily Value (PDV%). Based on our EDA analysis, calories in a recipe tends to increase as total fat content increases indicating there is some correlation between total fat and calories.
  2. Engineered Features:

    • protein_to_fat_ratio: This feature captures the ratio of protein to fat content in a recipe. This is a good engineered feature since fat contains more calories per gram (9 cal/g) compared to protein (4 cal/g). Therefore, a higher protein-to-fat ratio often correlates with a lower overall calorie density, helping to predict calorie content more accurately.

    • carbs_to_fat_ratio: This feature captures the ratio of carbohydrates to fat content in a recipe. This is a good engineered feature since fat contains more calories per gram (9 cal/g) compared to carbohydrates (4 cal/g). Therefore, a higher carbohydrates-to-fat ratio often correlates with a lower overall calorie density, helping to predict calorie content more accurately.

    • log_sugar The log transformation addresses skewness in sugar distribution, making relationships between sugar and calorie content more linear and manageable for regression.

    These features likely improve model performance as they help to capture more complex relationships between macronutrient compositions and the target variable.

  3. Response Variable:
    • calories: Number of calories, the response variable, measured as a continuous quantitative value.

Preprocessing

Modeling Algorithm

Best Hyperparameters

Through GridSearchCV, the optimal hyperparameter configuration was found to be:

The final model uses the engineered features and Lasso with cross-validation to identify optimal hyperparameters. To tune our Lasso regression model, we used GridSearchCV with 5-fold cross-validation to find the optimal value of alpha, the regularization strength. The search spanned alpha values from 0.1 to 100, and the best-performing value was 0.1, which offered a strong balance between model complexity and overfitting control. This approach ensured the model generalized well to unseen data by minimizing the average MSE across validation folds.

Model Performance

  1. Mean Squared Error (MSE):
    • Train MSE: 1560.999
    • Test MSE: 1323.506

    Interpretation: The train and test MSE values indicate the average squared difference between the predicted and actual calorie values. The relatively low MSE on both the training and test sets indicate that the model’s predictions are more accurate than the baseline and that it generalizes well to unseen data.

  2. R² Score (Coefficient of Determination):
    • Train R²: 0.9894
    • Test R²: 0.9911

    Interpretation: The model explains only ~99% of the variance in calorie values. This high degree of variance explanation suggests that the model captures the underlying relationships between the input features and calorie content effectively.

Final Model Improvement from Baseline

The baseline model was a simple linear regression model without feature engineering and parameter tuning. The improvements in the model are due to the increased complexity of the final model, the addition of features, feature engineering, and multiple iterations to find the best hyperparameters. By comparing mean squared errors (MSE) and R² scores between the training and test sets for both models, the final model shows a decrease in MSE and an increase in R², indicating better fit and generalization to unseen data.

Visualization of Model Performance

Final Project for UMich EECS 398 WN 25 using Recipes and Ratings Dataset