by Tatiana Samokhvalova
This project was created as an assignment for UCSD's DSC 80 class in Winter 2024.
In this project, I dive into an extensive dataset comprising detailed information on a wide variety of recipes and user interactions with these recipes. This dataset contains a variety of factors such as preparation time, nutritional values, step-by-step instructions, user reviews, and other. My investigation centeres around the relationship between the recipes' different features and their preparation times.
The 2 datasets I merged were interactions and recipes. The interactions dataset contains the user' reviews of different recipes and has 731927 rows. The recipes dataset contains information about different recipes and consists of 83782 rows. Here are the columns of both datasets:
user_id: id of the user who submitted a review.recipe_id: recipes's id (was used to merge the two datasets).date: the date on which the review was submitted.rating: the rating that was given to a recipe.review: text review of a recipe.
id: recipe's id.minutes: how long a recipe takes to make (recipe's preparation time).contributor_id: the id of the person who submitted the recipe.submitted: the date when the recipe was submitted.tags: different tags associated with a recipe.steps: the recipe's steps.nutrition: recipe's nutritional values.n_steps: the number of steps in the recipe.description: the recipe's text description.ingredients: the ingredients required for the recipe.n_ingredients: the number of ingredients.
As my goal was to predict recipes prepaation time, the relevant columns here would be minutes, n_steps, n_ingredients, and possibly description and some nutritional values.
First, as mentioned earlier, I replaced 0s in the rating column (before merging). I have come to that decision after seeing several positive reviews, or just questions and comments where the rating was 0. Hence, the value is not actually 0 and is in fact a missing value.
The recipes were merged on recipe_id, and I used the rating's average. Before the merge, I replaced 0s in the rating with NaN values, as they were identified as missing values and would affect the mean values (lower them). Furthermore, after merging I rounded the ratings to the nearest 0.5 so the variable would be categorical ordinal rather than continuous.
Then, I separated the column nutrition into 7 columns (calories, fat, sugar, sodium, protein, sat_fat, carbohydrates) for further analysis.
As a last step in this section, I dropped the following columns: nutrition (as it was transformed and the column itself became redundant), steps, ingredients, and tags (as I was not intending to use those columns for my analysis).
- The columns
stepsandindredientsare a free-form description of what a recipe is, and would not be useful for the model I planned on building. - As for the column
tags, I did not consider it to be useful for the overall analysis, and as my model's goal was predicting how long a meal would take to cook, it seemed unreasonable to use tags for it (as they appear on the recipe after it has been added to the website and are most likely the last thing to be added to a recipe).
This is what my dataframe looked after cleaning (only some columns are shown for easier reading):
| id | minutes | description | n_steps | n_ingredients | calories |
|---|---|---|---|---|---|
| 333281 | 40 | these are the most; chocolatey, moist... | 10 | 9 | 138.4 |
| 453467 | 45 | this is the recipe that we use at my... | 12 | 11 | 595.1 |
| 306168 | 40 | since there are already 411 recipes... | 6 | 9 | 194.8 |
| 286009 | 120 | why a millionaire pound cake? ... | 7 | 7 | 878.3 |
| 475785 | 90 | ready, set, cook! special ... | 17 | 13 | 267 |
I performed a univariate analysis on the rating column (that contained recipes' average ratings), and found that the ratings in the dataframe are heavily skewed. This means that the feature would be useless for analysis or making of the prediction model.
I examined the relationship between the recipe's number of ingredients and its calories. I found it interesting that at first, as expected, the amount of calories rises with the number of ingridients. However, for larger quantities of ingridients the amount of calories becomes lower on average.
<iframe src="ingr-calories-plot.html" width="600" height="600" frameborder="0" ></iframe>| Short | Medium | Long | Very long |
|---|---|---|---|
| 329.896 | 413.557 | 493.387 | 567.763 |
| 420.423 | 486.817 | 609.569 | 682.428 |
| 563.026 | 666.303 | 735.745 | 2472.27 |
| nan | 670.275 | 2180.86 | 559 |
This is a pivot table containing number of ingredients againt preparation times, with mean calories as values. However, first I will put n_ingredients into bins, as number of ingredients goes all the way up to 35. The table suggests calorie count rises with more ingredients but varies by preparation time; 'Very long' prep often has fewer calories, but that could also be attributed to sample size.
In the dataset, 3 columns showed missing data: name (MCAR), description (MAR), and rating (NMAR).
The rating column's missing mechanism is most likely NMAR, as people subjectively chose not to rate a recipe for different reasons: for example, when they want to ask a question or make a general comment without giving a recipe a grade. Some data that could be used as a confirmation here would be for example semantically analyzing people's reviews.
I have noticed that descriptions tend to be missing when recipes are simpler. So, description's missingness could be MAR-dependent on the column n-ingredients. I performed a permutation test and confirmed the dependency.
| n_ingredients_cat |
|---|
| 58 |
| 12 |
| 0 |
| 0 |
The question explored: Is there a strong correlation between the recipe's number of steps and its number of ingredients?
Null hypothesis: There is no relationship between n_steps and n_ingredients in the recipes.
Alternative hypothesis: There is a relationship between the number of calories and the amount of protein in the recipes.
Test stat: correlation between n_steps and n_ingredients.
After performing the test, with p-value < 0.05, I rejected the null hypothesis in facor of the alternative: there most likely is a correlation between a recipe's number of steps and its number of ingredients.
My prediction problem is to predict what preparation time category a recipe falls into ('Short', 'Medium', 'Long', 'Very long'). This is a multiclass classification problem. The features considered included mainly n_steps, n_ingredients or n_ingredients_cat, with the addition of calories and descriptions' containing certain keywords (such as 'quick'). The response variable is preparation_time (a categorical variable created from minutes). The metric used to validate the model is accuracy.
The features described are reasonable to have at the time of prediction, as I want to predict how long the recipe would take when it's already created, while such features as tags as most likely to be created after people have already cooked the recipes, hence the preparation time is already known.
The model uses a Decision Tree classifier with a maximum depth of 4 and three features: n_steps (quantitative) and n_ingredients_cat (ordinal, one-hot encoded). The model's accuracy is around 51%, which is technically better than randomly guessing (as there are 4 classes, so the chance to guess would be 25%), but there is a lot of room for improvement.
Features added include n_steps multiplied by n_ingredients, which reflects the interaction between recipe complexity and ingredient variety (potentially indicative of preparation time). Other features included are keywords in descriptions (like 'quick' or 'easy'), which hint at faster-to-prepare recipes, and recipes' calories, which provide a nutritional perspective, which could correlate with recipe complexity and preparation time. A Decision Tree classifier was chosen for its interpretability and simplicity, and was fitted using grid search. The best performing hyperparameters were a max_depth of 7, min_samples_leaf of 2, and min_samples_split of 50, identified through grid search and cross-validation (5-fold).
The Final Model shows a slight improvement over the Baseline (~1%, which is minimal improvement), suggesting the added features provide marginal additional predictive power, aligning with expectations about the data generating process and preparation time. However, even with added features and tuning of hyperparameters the model does not perform much better.
As the model is still performing poorly, I consider it pointless to perform fairness analysis as of this moment. However, in the next steps of my project, I will consider seeing how fair the model predicts preparation times for recipes from different decades, and how fair it performs on recipes with different nurtitional values.