Do Higher-Calorie Recipes Taste Better? 🍽️

Investigating the Relationship Between Calories and Ratings on Food.com

By: Tijil Chhabra

A DSC 80 Final Project at UC San Diego

Introduction

This project investigates the Recipes and Ratings dataset from food.com, originally scraped by Bodhisattwa Prasad Majumder et al. for their recommender systems research. The dataset contains recipes and user reviews posted since 2008.

Central Question: Do recipes with higher calorie counts tend to receive different average ratings than recipes with lower calorie counts?

This question is relevant because it explores whether indulgent, calorie-dense recipes genuinely earn better reviews from users or if lighter, healthier recipes are just as well-received. Understanding this relationship can help recipe developers, food bloggers, and health-conscious cooks make informed decisions.

The raw data comes in two CSV files:

Key columns relevant to our analysis:

Column Description
name Recipe name
id Recipe ID
minutes Minutes to prepare the recipe
n_steps Number of steps in the recipe
n_ingredients Number of ingredients in the recipe
nutrition Nutrition info as a string: [calories (#), total fat (PDV), sugar (PDV), sodium (PDV), protein (PDV), saturated fat (PDV), carbohydrates (PDV)]
rating Rating given by user (1–5), from the interactions dataset
description User-provided description of the recipe

Data Cleaning and Exploratory Data Analysis

Cleaning Steps

  1. Left-merged recipes with interactions on id/recipe_id so every recipe retains its row even if it has zero reviews. The merged dataset has 234,429 rows.
  2. Replaced 0 ratings with NaN β€” on food.com, the lowest possible rating is 1 star. A rating of 0 means the user wrote a review but chose not to leave a star rating, so it represents missing data, not an actual score.
  3. Computed average rating per recipe by grouping by recipe id and averaging non-null ratings, giving one aggregate score per recipe.
  4. Parsed the nutrition column from a string (e.g. '[138.4, 10.0, 50.0, ...]') into seven individual numeric columns: calories, total_fat_pdv, sugar_pdv, sodium_pdv, protein_pdv, saturated_fat_pdv, carbohydrates_pdv.
  5. Created calorie_category β€” High Calorie (above median of 305.4 calories) or Low Calorie β€” for use in hypothesis testing.
  6. Created rating_category β€” High (avg_rating > 4) or Low (avg_rating ≀ 4) β€” as the response variable for the classification task. 2,609 recipes have missing avg_rating (no reviews at all).

Here are the first few rows of the cleaned dataset:

name id minutes n_steps n_ingredients calories avg_rating calorie_category rating_category
1 brownies in the world best ever 333281 40 10 9 138.4 4.0 Low Calorie Low
1 in canada chocolate chip cookies 453467 45 12 11 595.1 5.0 High Calorie High
412 broccoli casserole 306168 40 6 9 194.8 5.0 Low Calorie High
millionaire pound cake 286009 120 7 7 878.3 5.0 High Calorie High
2000 meatloaf 475785 90 17 13 267.0 5.0 Low Calorie High

Univariate Analysis

The distribution of average ratings is heavily left-skewed, with the vast majority of recipes rated near 5.0. This is typical of voluntary review systems β€” users who are satisfied are more likely to leave a rating, while those who are indifferent or mildly dissatisfied may not rate at all.

The calorie distribution is right-skewed: most recipes are below 500 calories, with a long tail of high-calorie outliers. This is why we used a median-based split for grouping recipes into High and Low calorie categories.

Bivariate Analysis

Both calorie groups have similar median ratings (close to 5.0), but the High Calorie group shows slightly more spread in the lower quartiles β€” a subtle difference worth testing formally.

The scatter plot shows no strong linear trend between calories and rating. Recipes across the entire calorie spectrum cluster near 4–5 stars, consistent with the left-skewed rating distribution.

Interesting Aggregates

calorie_bin 1–5 steps 6–10 steps 11–20 steps 20+ steps
0–100 4.6697 4.6278 4.6316 4.6140
101–250 4.6461 4.6189 4.6191 4.6663
251–500 4.6256 4.6124 4.6187 4.6293
501–1000 4.5920 4.6128 4.6366 4.6641
1000+ 4.6203 4.6143 4.6123 4.6382

Average ratings are remarkably stable (~4.6–4.7) across all calorie ranges and step counts. Recipes with 501–1000 calories and 1–5 steps show a slight dip to 4.59, but overall, neither calorie count nor complexity has a dramatic impact on ratings.

Assessment of Missingness

MNAR Analysis

The rating column in the interactions dataset is likely MNAR (Not Missing At Random). On food.com, users can submit text reviews without leaving a star rating β€” in the raw data these appear as 0, which we replaced with NaN. The decision to skip the star rating likely depends on the rating value the user would have given: users with neutral or lukewarm opinions may not feel strongly enough to assign a number, while users with very strong positive or negative feelings are more motivated to rate. Since the probability of missingness depends on the unobserved value itself, this is MNAR.

To potentially make this MAR, we could collect additional data such as whether the user actually cooked the recipe, time spent on the recipe page, or whether the user saved/bookmarked the recipe. If the missingness could be fully explained by such observable variables, it would become MAR.

Missingness Dependency

Test 1: Does missingness of avg_rating depend on n_steps?

We performed a permutation test with 1,000 iterations, using the absolute difference in mean n_steps as the test statistic and a significance level of Ξ± = 0.05. The observed difference was 1.4930.

Result: The p-value is 0.0, so we reject the null hypothesis. The missingness of avg_rating depends on n_steps. Recipes with more steps tend to have different rates of missing ratings β€” likely because complex recipes attract fewer total reviews.

Test 2: Does missingness of avg_rating depend on sodium_pdv?

We performed a permutation test with 1,000 iterations using the Kolmogorov-Smirnov (K-S) statistic β€” the maximum absolute difference between the two groups’ CDFs. We chose the K-S statistic over difference in means because with 83,000+ rows, even trivially small differences in means can appear statistically significant; the K-S statistic better measures whether the two distributions actually differ in shape. The observed K-S statistic was 0.0237.

Result: The p-value is 0.087, which is greater than Ξ± = 0.05, so we fail to reject the null hypothesis. The missingness of avg_rating does not appear to depend on sodium_pdv. A recipe’s sodium content does not significantly predict whether it will have missing ratings.

Hypothesis Testing

Question: Do recipes with higher calorie counts have different average ratings than lower-calorie recipes?

We split recipes into two groups at the median calorie count (305.4 calories):

Null Hypothesis: Recipes with higher calorie counts and recipes with lower calorie counts have the same average rating. Any observed difference is due to random chance.

Alternative Hypothesis: Recipes with higher calorie counts have a different average rating than recipes with lower calorie counts.

Test Statistic: Difference in group means (High Calorie βˆ’ Low Calorie).

Significance Level: Ξ± = 0.05

We chose a permutation test because we are comparing two observed groups from the same dataset. The difference in means is natural for comparing a quantitative variable across two groups, and a two-sided test is appropriate since we have no prior directional expectation.

Observed results:

Result: The p-value is 0.0721, which is greater than Ξ± = 0.05, so we fail to reject the null hypothesis. We do not have sufficient evidence to conclude that calorie content is associated with a difference in average ratings. The tiny observed difference of βˆ’0.008 rating points is consistent with random variation. This is an observational study, so even if we had rejected the null, we could not have concluded causation.

Framing a Prediction Problem

Problem: Predict whether a recipe will receive a High (avg_rating > 4) or Low (avg_rating ≀ 4) average rating.

Type: Binary classification.

Response Variable: rating_category β€” chosen because the raw avg_rating distribution is extremely left-skewed (most recipes are ~5 stars), making exact numeric prediction very difficult. Binarizing at 4 creates a more balanced and interpretable task.

Metric: F1-score (weighted), chosen over accuracy because the classes are imbalanced β€” 63,606 High vs. 17,567 Low (78% vs. 22%). A model that always predicts β€œHigh” would achieve 78% accuracy but be useless. F1-score penalizes models that ignore the minority class.

Features at time of prediction: Only recipe attributes known before any ratings exist: calories, n_steps, n_ingredients, minutes, sugar_pdv, protein_pdv, sodium_pdv, total_fat_pdv, saturated_fat_pdv, carbohydrates_pdv. We do NOT use rating, avg_rating, review, or date.

Baseline Model

Model: Decision Tree Classifier (default hyperparameters).

Features (2):

  1. n_ingredients β€” quantitative. Reflects recipe complexity. Left as-is (passthrough) since decision trees handle numeric features natively.
  2. calorie_category β€” nominal (High Calorie / Low Calorie). Encoded with OneHotEncoder(drop='if_binary'), creating a single binary column.

Performance:

This baseline model is not good. While it achieves 78% accuracy, this is misleading β€” the model predicts β€œHigh” for nearly every recipe and completely ignores the β€œLow” class (0.00 recall, 0.00 F1 for β€œLow”). This is a consequence of class imbalance: about 78% of recipes are rated above 4, so a naive model always predicting β€œHigh” achieves similar accuracy. The identical training and test scores suggest the model is underfitting β€” with only two features, it cannot learn meaningful patterns.

Final Model

Algorithm: Random Forest Classifier with class_weight='balanced'.

New Features Added (beyond baseline):

Hyperparameter Tuning: GridSearchCV with 5-fold cross-validation over max_depth ∈ {5, 10, 20}, min_samples_split ∈ {10, 50, 100}, n_estimators ∈ {100, 200}.

Best Hyperparameters: max_depth=20, min_samples_split=10, n_estimators=200.

Performance:

The key improvement was adding class_weight='balanced', which forces the Random Forest to penalize misclassification of β€œLow” recipes more heavily. This dramatically improved Low recall from 0% to 12% and Low F1 from 0.00 to 0.16. While the overall improvement in weighted F1 is modest (+0.0046), the model now meaningfully identifies some low-rated recipes instead of blindly predicting β€œHigh” for everything. This is a fundamentally difficult prediction task β€” the vast majority of recipes on food.com are rated above 4 stars, leaving limited signal to distinguish the minority class from recipe attributes alone.

Fairness Analysis

Question: Does our model perform equally well for high-calorie vs. low-calorie recipes?

Observed Results:

Result: The p-value is 0.0, which is less than Ξ± = 0.05, so we reject the null hypothesis. Our model appears to be unfair β€” it performs significantly differently for high-calorie recipes (F1 = 0.679) compared to low-calorie recipes (F1 = 0.706). This is not entirely surprising since calorie content is one of the model’s features, so the model may have learned different decision boundaries for each group. Additionally, the class balance (proportion of High vs. Low rated recipes) may differ between the two calorie groups, contributing to unequal performance.