By: Aayush Sharma
Contact: aayushsh@umich.edu
This project explores recipe and user interaction data from Food.com, a major recipe-sharing platform. The dataset includes thousands of recipes, their nutrition, user-submitted ratings, and reviews.
Investigative Question:
How can we predict the number of calories for a recipe?
Understanding how recipe characteristics influence calorie count is important for:
By modeling calorie content based on recipe features, we can gain insight into which types of recipes and preparation methods contribute most to total caloric value.
We use two datasets from Food.com.
RAW_recipes.csvDescription: Contains 83,782 recipes with ingredients, steps, and nutrition information.
Table 1: Sample from RAW_recipes.csv
| name | id | minutes | nutrition | n_steps |
|---|---|---|---|---|
| 1 brownies in the world best ever | 333281 | 40 | [138.4, 10.0, 50.0, 3.0, 3.0, 19.0, 6.0] | 10 |
| 1 in canada chocolate chip cookies | 453467 | 45 | [595.1, 46.0, 211.0, 22.0, 13.0, 51.0, 26.0] | 12 |
| 412 broccoli casserole | 306168 | 40 | [194.8, 20.0, 6.0, 32.0, 22.0, 36.0, 3.0] | 6 |
| millionaire pound cake | 286009 | 120 | [878.3, 63.0, 326.0, 13.0, 20.0, 123.0, 39.0] | 7 |
| 2000 meatloaf | 475785 | 90 | [267.0, 30.0, 12.0, 12.0, 29.0, 48.0, 2.0] | 17 |
RAW_interactions.csvDescription: Contains 731,927 user reviews and ratings.
Table 2: Sample from RAW_interactions.csv
| recipe_id | rating | date | user_id |
|---|---|---|---|
| 40893 | 5 | 2011-12-21 | 1293707 |
| 85009 | 5 | 2010-02-27 | 126440 |
| 85009 | 5 | 2011-10-01 | 57222 |
| 120345 | 0 | 2011-08-06 | 124416 |
| 120345 | 2 | 2015-05-10 | 2000192946 |
| Column | Description |
|---|---|
calories |
Number of calories per serving (target variable) |
total_fat |
Total fat content (% Daily Value) |
sugar |
Sugar content (% Daily Value) |
sodium |
Sodium content (% Daily Value) |
protein |
Protein content (% Daily Value) |
saturated_fat |
Saturated fat content (% Daily Value) |
n_ingredients |
Number of ingredients |
n_steps |
Number of preparation steps |
rating |
Rating between 1 - 5 |
These features will be used to build a model capable of predicting the calorie content of a recipe based on its nutritional profile and complexity.
To prepare the dataset for analysis, I carried out a series of data cleaning steps aimed at ensuring the quality and reliability of the information. Below is a breakdown of each step taken.
I began by merging the two datasets. I performed a left join using the recipe ID to combine these datasets, ensuring that all recipe entries were preserved even if some did not have associated user feedback. After the merge, I removed the redundant recipe_id column that was duplicated in the merged result.
Some recipes had a rating value of 0, which is not a valid rating in the context of the platform (ratings typically range from 1 to 5). These zeroes likely represented missing or incorrect data. To avoid skewing the results, I replaced all zero values in the rating column with NaN to mark them as missing.
I grouped the data by recipe ID and computed the mean of the valid (non-missing) ratings. This provided a more accurate picture of how each recipe was rated by users on average.
I merged the calculated average rating back into the main dataset so that each row, which represents a user interaction or review, also included the average rating for the corresponding recipe.
The dataset included a nutrition column that stored nutritional data as a stringified list. I parsed this list and split it into individual columns, including:
caloriestotal_fatsugarsodiumproteinsaturated_fatcarbohydratesOnce these values were extracted and structured, I removed the original nutrition column to eliminate redundancy.
After these cleaning steps, the dataset was significantly more structured, with invalid data removed and additional features such as average ratings and detailed nutrition values included to support deeper analysis.
| name | total_fat | carbohydrates | sodium | sugar | protein | rating_average | n_steps |
|---|---|---|---|---|---|---|---|
| 1 brownies in the world best ever | 10 | 6 | 3 | 50 | 3 | 4 | 10 |
| 1 in canada chocolate chip cookies | 46 | 26 | 22 | 211 | 13 | 5 | 12 |
| 412 broccoli casserole | 20 | 3 | 32 | 6 | 22 | 5 | 6 |
| 412 broccoli casserole | 20 | 3 | 32 | 6 | 22 | 5 | 6 |
| 412 broccoli casserole | 20 | 3 | 32 | 6 | 22 | 5 | 6 |
The violin plot above visualizes the distribution of calorie values across all recipes, with outliers removed for clarity. The shape of the plot highlights the density of data points — showing that most recipes cluster around the 175–450 calorie range. The thickest part of the distribution, near the median, gives a quick sense of central tendency, while the tails show a right skew on the data.
This visualization is particularly useful for the calorie prediction task ahead. It helps identify:
Understanding this distribution ensures the model is trained on realistic, well-ordered data.
This histogram shows the distribution of sugar content per recipe, with outliers removed. Most recipes contain less than 20% of the daily value of sugar, with a sharp peak around the very low end — possibly indicating savory dishes or low-sugar meals. There is a long right tail, which suggests a minority of recipes (e.g. desserts or drinks) with much higher sugar content.
Understanding this distribution is crucial for building a calorie prediction model. Since sugar is often a strong contributor to calorie count, observing its skew and overall spread helps in:
This insight can guide feature engineering pertaining to sugar for the final model.
These grouped boxplots show how calorie counts varies across different levels of total fat content, with outliers removed. There’s a clear upward trend — as total fat increases, so do the calories. The distribution also widens, indicating more variability in high-fat recipes. This supports the idea that total fat is a strong predictor of calories and will likely be an important feature in the calorie prediction model. The consistent increase in medians across bins shows a strong positive correlation between total fat and calorie count. Additionally, the widening interquartile ranges at higher fat levels suggest that calorie content becomes more dispersed, possibly due to the inclusion of richer, more energy-dense foods.
This plot directly addresses the question of whether fat content can help predict calories, and the visible trends strongly support that relationship.
| Steps Group | Calories | Protein(PDV) | Sugar(PDV) | Total Fat(PDV) |
|---|---|---|---|---|
| Low | 316.7 | 21.6 | 25.4 | 20.4 |
| Medium | 423.8 | 29.2 | 30.4 | 29.4 |
| High | 508.2 | 40.9 | 71.2 | 39.3 |
This pivot table groups recipes by the number of steps and shows the average nutritional values for each group. There’s a clear upward trend: recipes with more steps tend to have more calories, protein, sugar, and fat.
This is significant because it suggests that step count could be a useful feature when predicting calories. It captures a pattern in the data that could help a model better estimate calorie content based on how complex a recipe is.
The only imputation performed was shown in the Data Cleaning section.
In this context, imputing nutritional values (like calories, protein, sugar, etc.) isn’t appropriate — these features are specific to the ingredients in each recipe. Filling them with averages or estimates would distort the data and potentially mislead any predictive models. As a result, dropna() was used to drop any null rows.
Out of 234,429 original recipes, 15,198 were removed due to missing values — a 6.48% decrease in dataset size. This was an acceptable trade-off to preserve data integrity without introducing noise.
How can we predict the number of calories in a recipe?
Regression
This is a regression problem because the target variable, calories, is a continuous numerical value. The goal is to predict an exact number rather than assign the input to a predefined category. Since the output can span a wide range of values, this aligns with the nature of regression tasks. If the target were categorical, such as classifying recipes into calorie ranges, it would be a classification problem.
The target variable in this prediction task is calories, representing the total caloric content of a recipe.
Calories are a fundamental measure of a recipe’s nutritional profile. Estimating calories provides value in several ways:
The model is evaluated using Mean Squared Error (MSE).
MSE is chosen because it:
For the baseline regression model, I used fat and carbohydrates as baseline features because they’re key macronutrients that directly impact a food item’s nutritional value, especially calorie content. Since both are numerical, they required no encoding and were a simple, effective starting point for evaluating model performance.
fat, carbohydratesI didn’t perform any encodings since all selected features were already numerical.
The entire modeling process was implemented using sklearn’s Pipeline to keep preprocessing and model training streamlined. The model used was LinearRegression.
Model performance was evaluated using Mean Squared Error (MSE):
8,906.0898,915.76810,274.837The training and validation errors are very close, suggesting that the model generalizes reasonably well to unseen data and is not overfitting. However, the relatively high error values indicate that there’s room for improvement — likely by incorporating more informative features.
The final model uses a RandomForestRegressor inside a Pipeline, optimized with randomized hyperparameter search. The pipeline includes both preprocessing and modeling steps to ensure consistency and reproducibility.
Total: 10 features
total_fatcarbohydratessodiumsugarproteinfat_carb_ratio (engineered)macro_total (engineered)protein_ratio (engineered)n_stepsrating (one-hot encoded)total_fat: Major contributor to calories (9 cal/g); normalized due to skew.carbohydrates: Key macronutrient (4 cal/g); transformed to stabilize variance.sodium: Correlated with processed foods and caloric density; transformed to handle outliers.sugar: High sugar usually implies high calories; normalized for consistency.protein: Another caloric macro (4 cal/g); normalized to align scales.fat_carb_ratio: Captures balance between two major macros and can give insight on calorie density.macro_total: Sum of all macros; strong linear signal for calories.protein_ratio: Measures how protein-dense a food is relative to other macronutrients. Including this feature helps the model distinguish between lean/high-protein foods and carb- or fat-dominant ones, even if total protein content is similar. It introduces a relative perspective that improves the model’s understanding of food composition.n_steps: Indicates how complex a recipe is to prepare. Recipes with more steps are often richer or more elaborate (e.g., multi-layered meals, sauces), which tend to be higher in calories. Including this feature provides context about the likely richness or simplicity of a dish that raw nutritional values don’t capture.rating: A user-generated rating that can reflect how enjoyable or satisfying a food is. Highly rated foods are often indulgent or rich, which can align with higher calorie content. By including this feature, the model incorporates human preference signals that may correlate with calorie density.QuantileTransformer with n_quantiles=50, mapping skewed distributions to a normal shape.rating) was encoded using OneHotEncoder with the first category dropped to prevent redundancy.To improve model performance and prevent overfitting, I used RandomizedSearchCV to search for the best combination of hyperparameters for the RandomForestRegressor. This method randomly samples a fixed number of combinations from a predefined grid of possible values, rather than exhaustively evaluating all options like GridSearchCV. This makes it much more efficient, especially when the search space is large.
The randomized search was run over 20 different combinations, using 5-fold cross-validation to ensure reliable validation metrics and avoid depending on a single train/validation split.
Scoring metric:
Parameters that were tuned:
model__n_estimators: Number of trees in the forestmodel__max_depth: Maximum depth of each treemodel__min_samples_split: Minimum number of samples required to split a nodemodel__min_samples_leaf: Minimum number of samples required to be at a leaf nodemodel__max_features: Number of features considered when looking for the best splitpreprocessor__quantile_features__n_quantiles: Number of quantiles for the QuantileTransformer, which affects how distributions are normalizedAfter tuning, the best combination of parameters was selected based on the lowest validation MSE across the 5 folds.
Best Parameters Found:
preprocessor__quantile_features__n_quantiles: 50model__n_estimators: 200model__max_depth: 10model__min_samples_split: 5model__min_samples_leaf: 2model__max_features: None| Metric | Value |
|---|---|
| Training MSE | 1892.21 |
| Validation MSE | 5090.05 |
| Testing MSE | 8680.47 |
The final model demonstrates stronger performance, especially in reducing training and validation error compared to the baseline. However, the large gap between training and validations shows potential signs of overfitting.
Yes — the final model is a clear improvement over the baseline, especially where it matters most: on unseen data.
The baseline model had a test MSE of approximately 10,275, while the tuned final model achieved a significantly lower test MSE of around 8680.47. This drop of over 1,400 in error demonstrates that the final model generalizes better and makes more accurate predictions when faced with new data.
This improvement is due to several key changes:
macro_total and protein_ratio.rating and n_steps, which capture patterns that raw nutrition data may miss.RandomizedSearchCV to fine-tune model complexity.While training and validation MSEs also improved significantly, the ultimate goal is a model that performs well on new, real-world data — and this final model does just that.