Author: Keanu Ventura
“Every second counts”
— The Bear
In the stressful and exacting world of cooking, time, ingredients, preparation, and other factors all play a crucial role in determining the outcome of a dish. In one of my favorite TV series, The Bear, special emphasis is placed on the importance of time, not just in the kitchen, but in life as well. Inspired by this, my data science project dives into two large datasets consisting of recipes and ratings posted since 2008 on Food.com to answer one central question:
What is the relationship between the cooking time and the average rating of recipes?
The first dataset, RAW_recipes.csv, contains 83,782 rows, indicating 83,782 recipes, and 12 columns providing the following information:
| Column | Description |
|---|---|
| name | Recipe name |
| id | Recipe ID |
| minutes | Minutes to prepare recipe |
| contributor_id | User ID who submitted this recipe |
| submitted | Date recipe was submitted |
| tags | Food.com tags for recipe |
| nutrition | Nutrition info: calories, fat, sugar, sodium, protein, saturated fat, carbs (% daily value) |
| n_steps | Number of steps in recipe |
| steps | Text for recipe steps, in order |
| description | User-provided description |
| ingredients | Text for recipe ingredients |
| n_ingredients | Number of ingredients in recipe |
The second dataset, RAW_interactions.csv, contains 731,927 rows, indicating 731,927 reviews from users on recipes, and 5 columns providing the following information:
| Column | Description |
|---|---|
| user_id | User ID |
| recipe_id | Recipe ID |
| date | Date of interaction |
| rating | Rating given |
| review | Review text |
Understanding what makes a recipe have highly rated enables home cooks, food enthusiasts, and recipe developers to make informed decisions when selecting or creating meals. By investigating the relationship between time and ratings, we can discover if there are patterns that illustrate what people value in a recipe. This project will check if people think spending more time leads to better quality recipes or if they just want quick and easy meals.
“I need you to understand that you’re not just cleaning, you’re resetting the systems”
— Sydney, The Bear
The following data cleaning steps I took were:
The final dataset, cleaned, contains 218885 rows, indicating the recipe with their respective review(s) retained after cleaning, and 19 columns with the following data types:
| Column | Data Type |
|---|---|
| name | object |
| id | int64 |
| minutes | int64 |
| contributor_id | int64 |
| submitted | datetime64[ns] |
| tags | object |
| nutrition | object |
| n_steps | int64 |
| steps | object |
| description | object |
| ingredients | object |
| n_ingredients | int64 |
| user_id | float64 |
| date | datetime64[ns] |
| rating | float64 |
| review | object |
| avg_rating | float64 |
| time_bins | category |
| calories | float64 |
Here are a few rows of cleaned, showing columns relevant to the future analysis:
| name | minutes | n_steps | n_ingredients | rating | avg_rating | time_bins | calories |
|---|---|---|---|---|---|---|---|
| 1 brownies in the world best ever | 40 | 10 | 9 | 4 | 4 | 40-50 | 138.4 |
| 1 in canada chocolate chip cookies | 45 | 12 | 11 | 5 | 5 | 40-50 | 595.1 |
| 412 broccoli casserole | 40 | 6 | 9 | 5 | 5 | 40-50 | 194.8 |
For this analysis, I analyzed the distribution of cooking time (minutes) for the recipes. The histogram below depicts how frequently different cooking times appeared in the dataset. The distribution is skewed to the right with a long right tail. There is also a decreasing trend, indicating that as cooking time gets longer, the amount of recipes on Food.com gets lower.
For this analysis, I analyzed the average recipe ratings across different durations. The line graph below illustrates the average rating for each cooking time interval. The distribution shows a clear downward trend, revealing that as cooking time increases, the average ratings tend to slightly decrease. This suggests that recipes that take longer to make may not be rated as highly as those with shorter times.
To further understand the relationship between cooking time and average rating, I computed summary statistics for each time bin. Looking at the mean and median average ratings, the values remain consistently high across the bins, however, there is a noticeable decline once it reaches the 240-300 minute range. This suggests that once recipes exceed 240 minutes, their reviews have lower ratings. Additionally, the 0-20 minutes bins present the highest mean and median average ratings, suggesting that people may favor quick recipes. While every time interval contains recipes rated from 1-5, it is evident that recipes with shorter cooking durations are more frequent in the dataset, indicating that fast meals are not only more loved on Food.com, but more common.
| time_bins | mean_avg_rating | median_avg_rating |
|---|---|---|
| 0-10 | 4.71089 | 4.9 |
| 10-20 | 4.71461 | 4.90909 |
| 20-30 | 4.67489 | 4.85714 |
| 30-40 | 4.67723 | 4.85714 |
| 40-50 | 4.66288 | 4.85 |
| 50-60 | 4.66885 | 4.85185 |
| 60-90 | 4.67527 | 4.88095 |
| 90-120 | 4.66838 | 4.9 |
| 120-180 | 4.66414 | 4.875 |
| 180-240 | 4.66297 | 4.83333 |
| 240-300 | 4.58725 | 4.69231 |
| 300-400 | 4.5509 | 4.67742 |
| 400-500 | 4.57296 | 4.73333 |
| time_bins | min_avg_rating | max_avg_rating | count_avg_rating |
|---|---|---|---|
| 0-10 | 1 | 5 | 21011 |
| 10-20 | 1 | 5 | 30721 |
| 20-30 | 1 | 5 | 35736 |
| 30-40 | 1 | 5 | 35190 |
| 40-50 | 1 | 5 | 25941 |
| 50-60 | 1 | 5 | 17356 |
| 60-90 | 1 | 5 | 32605 |
| 90-120 | 1 | 5 | 7985 |
| 120-180 | 1 | 5 | 7770 |
| 180-240 | 1 | 5 | 4114 |
| 240-300 | 1 | 5 | 3372 |
| 300-400 | 1 | 5 | 3913 |
| 400-500 | 1 | 5 | 2310 |
“This is a system. It’s not just chaos. Every little thing matters”
— Carmy, The Bear
I believe that the rating column in my data set is Not Missing At Random (NMAR) because the missingness of a rating may depend on the user_id column. This is because certain users who leave reviews on Food.com may have a habit of consistently choosing not to leave a rating after leaving a review. Users’ individual behavior may be linked to the missingness and is not captured by any of the other observed variables. Thus to explore this, I performed the following permutation test:
Null Hypothesis (H₀): The missingness of ratings does not depend on the user_id
Alternative Hypothesis (H₁): The missingness of ratings does depend on the user_id
Test Statistic: The variance of the proportions of missing ratings across different user_id’s
Significance Level: 5% (α = 0.05)
The bar chart below displays the top 20 user_id’s that have the most missing ratings, highlighting how some users consistently leave reviews on recipes without leaving a rating. The distribution supports the notion that the missingness is user-dependent.
The plot below displays the empirical distribution of the variance in missing ratings across user id’s under the null hypothesis, with the observed statistic of 0.128 marked. This visualization highlights how extreme the observed value is compared to what we’d expect by random chance, providing strong evidence against the null.
Results: The observed variance of the proportions of missing ratings across different user_id’s was 0.128, and the resulting p-value was 0.0
Conclusion: Since the p-value of 0.0 is less than the significance level of 0.05, we reject the null hypothesis. This suggests there is strong evidence that the missingness of ratings does depend on the user_id, supporting the idea that the data is NMAR
Building on the previous test that proved missingness depends on user_id, I also tested whether the missingness of the rating columns depends on the minutes column. I did this to prove that missingness is not related to cooking time and to reinforce that the missingness pattern is specifically tied to user behavior and not random concerning other variables. Thus, to explore this, I performed the following permutation test:
Null Hypothesis (H₀): The missingness of ratings does not depend on the cooking time of the recipe in minutes
Alternative Hypothesis (H₁): The missingness of ratings does depend on the cooking time of the recipe in minutes
Test Statistic: The variance of the proportions of missing ratings across different cooking times
Significance Level: 5% (α = 0.05)
The histogram below portrays the distribution of cooking time by rating missingness, highlighting that missing and non-missing rating values are similarly distributed across cooking times. This suggests that the missingness does not depend on the minutes column.
The plot below shows the empirical distribution of the variance in missing ratings across cooking times under the null hypothesis, with the observed statistic of 0.0131 marked. Since the observed value falls well within the simulated distribution, it supports the conclusion that the missingness of ratings is unrelated to cooking time.
Results: The observed variance of the proportions of missing ratings across different cooking times was 0.0131, and the resulting p-value was 0.297
Conclusion: Since the p-value of 0.297 is greater than the significance level of 0.05, we fail to reject the null hypothesis. This suggests there is strong evidence that the missingness of ratings does not depend on the cooking time (minutes)
“I just thought it’d be nice… something simple, something good”
— Carmy, The Bear
To investigate the relationship between cooking time and recipe ratings, I performed a hypothesis test comparing recipes with short cooking times (20–30 minutes) to those with long cooking times (300–400 minutes)
Null Hypothesis (H₀): There is no difference in average recipe ratings between recipes with short cooking times and those with long cooking times
Alternative Hypothesis (H₁): There is a difference in average recipe ratings between the two cooking time bins
Test Statistic: The difference in means of average ratings between the two cooking time bins
Significance Level: 5% (α = 0.05)
The histogram below shows the distribution of differences in average ratings between the 20–30 minute and 300–400 minute cooking time bins, with the observed statistic of 0.124 marked. The observed difference lies far outside the permuted differences, providing strong evidence against the null hypothesis.
Results: The observed difference in average ratings was approximately 0.124, and the resulting p-value was 0.0
Conclusion: Since the p-value of 0.0 is less than the significance level of 0.05, we reject the null hypothesis. This suggests there is strong evidence that recipes with short and long cooking times tend to receive different average ratings
“What grows together, goes together”
— Tina, The Bear
For my project, I plan to predict the number of calories in a recipe.
Type of problem: Regression
Response variable: Calories (continuous)
Evaluation Metric: R²
Time of prediction information: For the features I would know at the time of prediction, I decided to stray away from features like the full nutrition breakdown (e.g., fat, sugar) as they would only be known after calories are already calculated. Thus, I selected features known before the nutritional information was calculated such as n_ingredients and minutes
“Start from scratch. Make it clean. Make it simple”
— Marcus, The Bear
For my baseline model, I built a Linear Regression model using two features I believed would have the best correlation with calories: n_ingredients and minutes. I inferred that the more ingredients in a recipe and the longer it takes to cook a recipe, the more calories it may contain. Additionally, for my model, I constructed a pipeline using a Column Transformer to pass these features directly into the Linear Regression model. After fitting the model to the cleaned dataset, I calculated the evaluation metric, R².
Results: The R² was 0.0856
Conclusion: This means that my model explains around 8.56% of the variance in calorie content. The result suggests that my baseline model does not fully capture the complexity in the data and there is much to be done to advance my model and improve its performance
“This is not a sprint. It’s a marathon, and we’re learning every single day”
— Carmy, The Bear
For my final model, I first added n_steps alongside n_ingredients and minutes to expand the list of features for feature engineering. I chose n_steps specifically because I believed that a recipe with more steps may have more calories as longer and more complex recipes often involve additional ingredients and techniques that can contribute to a higher calorie content. To better prepare the data, in my preprocessor, I used a Standard Scaler to scale n_ingredients and n_steps and applied a Quantile Transformer to minutes to normalize its distribution. Additionally, for my final model, I used a Random Forest Regressor with hyperparameter tuning on tree depth and number of estimators via 5-fold cross-validation.
Results: The best parameters found using grid search were ‘regressor_max_depth’: 5 and ‘regressor_n_estimators’: 1000. The R² was 0.1087
Conclusion: According to these results, my final model explains around 10.87% of the variance in calorie content. This suggests that compared to my baseline model, although the improvement is small, my final model shows an improved R² and explains more variance in calorie counts. Thus, taking into account n_steps, better preparing the data, using a Random Forest, and using hyperparameter tuning added more complexity to my model. However, my outcome highlights the difficulty of predicting a recipe’s calories using the given data
“You don’t get to decide how people feel”
— Sydney, The Bear
To assess fairness in my model, I compared its performance across two groups:
Group X: Simple recipes (n_steps ≤ 9)
Group Y: Complex recipes (n_steps > 9)
Evaluation metric: Root mean squared error (RMSE)
Null Hypothesis (H₀): My model is fair. Its RMSE for simple and complex recipes is roughly the same, and any differences are due to random chance
Alternative Hypothesis (H₁): My model is unfair. Its RMSE for simple recipes is lower than that for complex recipes
The histogram below displays the distribution of RMSE differences between complex and simple recipes, with the observed difference of 10.785 marked. The observed difference lies far outside the range of typical permutation values, indicating that there is unfairness in my model’s performance across the two groups.
Results: The observed RMSE for simple recipes was 197.1212, while the observed RMSE for complex recipes was 207.9062, resulting in an observed difference of 10.785. The resulting p-value was 0.0.
Conclusion: Since the p-value of 0.0 is less than the significance level of 0.05, we reject the null hypothesis. This provides strong evidence that my model is unfair, as it performs significantly worse on complex recipes than on simple ones. In other words, the model is less accurate when predicting calorie content for recipes with larger n_steps.