The Relationship Between Calories and Recipe Rating

Introduction

This data science project, conducted at UCSD, explores the relationship between calories and average recipe ratings. The complete analysis and source code are available on GitHub. According to the Food and Drug Administration (FDA), calories are a measure of how much energy you get from one serving of food. Consuming more calories than you use on a daily basis is linked to overweight and obesity. The dataset used for this project contains recipes and ratings from food.com posted since 2008. This dataset was originally scraped and used in the paper, Generating Personalized Recipes from Historical User Preferences, by Majumder et al.

Using these datasets I explored whether the recipes with high calories were rated lower than recipes with low calories. The answer to this question could provide useful insight into how calories affect an individual's perception of a recipe.

Data Cleaning and Exploratory Data Analysis

To make the datasets more suitable for my analysis I made the following transformations to the recipes and ratings datasets.

1. Left merge the recipes and ratings datasets together.
2. Filled all ratings of 0 with np.nan in the merged dataset.
3. Calculated the average rating per recipe, as a Series.
4. Added this the average recipe rating Series back to the recipes dataset.
5. Converted nutrition, tags, steps, and ingredients columns from str to list.
6. Split the nutrition column into individual columns.
7. Added calorie_level column that labels recipes as either "High" or "Low" depending if their calories are greater than or less than the median for calories.
8. Drop columns unrelated to my topic.

Result

The cleaned dataset features 83,782 rows and 20 columns. Below are the selected columns and their respective data types.

Univariate Analysis

Bivariate Analysis

For my bivariate analysis, I created two columns is_dietary and is_healthy to help with the analysis. is_dietary indicates if a recipe contains the dietary tag and is_healthy indicates if a recipe conatains the healthy tag.

Interesting Aggregates

For this section, I investigated the relationship between the number of ingredients and calories of the recipes. First, I dropped outliers in the calories column using the IQR method. Then, I grouped the cleaned recipes dataframe by n_ingredients. I then selected the calories column and applied the aggregate functions, mean, median, and count.

Below is a visual representation of this table with two y-axis to view the data more clearly. It appears that the most popular amount of ingredients to have is 8. This shows that a majority of recipes have a moderate complexity in terms of ingredient count. Both the mean and median calorie counts tend to increase steadily as the number of ingredients increases. This suggests that more complex recipes generally contain higher calorie counts. The mean and median calories track closely together, indicating that the calorie distribution for each number of ingredients is roughly symmetric.

Assessment of Missingness

NMAR Analysis

I believe the review column from the ratings dataset is not missing at random (NMAR). People who viewed a recipe as mediocre or forgettable would not be inclined to leave a review because they would not have much to praise or criticize about the recipe. If the recipe was just in the middle of the road for them they would not feel strongly enough to take the time and effort to write and publish a review. In contrast, an individual who really enjoyed or despised the recipe would be much more inclined to put in the time and effort to leave a review for the recipe.

Missingness Dependency

I filtered the cleaned recipes dataset to see which column had the most missing values and found that avg_rating had the most missing values. I started by creating a avg_rating_missing column that indicated if the average rating was missing for the recipe.

Hypothesis Testing

For my hypothesis test, I explored the relationship between a recipe's average rating and the calorie level. As mentioned in data cleaning, recipes with a calorie count greater than the median were classified as "high calorie" and recipes with a calorie count below were classified as "low calorie." To analyze this relationship, I conducted a permutation test.

I decided to use a permutation test because it does not rely on distributional assumptions and is well suited for comparing groups in this scenario. I proposed that the average rating for low-calorie recipes is higher than higher-calorie recipes because people may be worried about the negative health risks present in the recipe. I chose the difference of means as the test statisitic because I wanted to find the whether low-calorie recipes were rated higher on average than high-calorie recipes. The difference of means, opposed to the absolute difference of means, shows the which group has a higher mean average rating.

I shuffled the avg_rating column 1000 times and calculated the mean differences between the low and high calorie groups.

Framing a Prediction Problem

I aim to predict a recipe's average score. I plan to use a classification model that classifies recipes into three categories, high, medium, and low. High would be scores greater than 3.75, medium would be scores between 3.75 and 2.75, and low would be any score less than 2.75. I opted to use a classification model opposed to a regression model because there is not much variance amongst the avg_rating column. We also found out earlier that the average rating distirbution is heavily skewed to the left, with most recipes receiving high ratings (primarily 5). To address this imbalance and to focus on predicting meaningful distinctions, I decided to classify the average ratings into high, medium, and low categories. This approach allows the model to better capture these distinct groups instead of trying to predict a nearly uniform continuous target. To evaluate this model, I will be using F1 score and accuracy because avg_rating is heavily imbalanced, with the majority of ratings being "high." The F1 score takes into account both precision and recall across all classes and gives more weight to the performance on the larger classes. Accuracy provides an overall snapshot of how many predictions were correct. Together, these metrics give a comprehensive picture of how well the model performs across all classes.

Baseline Model

For my baseline model I am implementing a random forest classifier. I will be using the features, calories and is_dietary to help classify avg_rating as high, medium, or low. First, I created a avg_rating_class column that indicated if a recipe was high, medium, ore low calorie. Then, I split the data points into training and testing sets to be used to train and test my model.

The F1 score of this model was 0.90 and the accuracy of this model was 0.92. These scores are really high and suggest the model is performing as intended. Although this may just be a facade, the F1 score for "high" is 0.96, but for "medium" and "low" the F1 score is a measly 0.01 for both. This is likely due to the fact that a majority of the average ratings falls above 4.

Final Model

For my final model I engineered new features that could help. These features were avg_step_desc_length, description_length, and year_submitted.

I also implemented GridSearchCV to hyperparameterize my random forest classifier. I aimed to tune the parameters, n_estimators, max_depth, and min_samples_split. It was found that the best parameters were a 50 n_estimators, a max depth of 5, and a min_samples_split of 2.

My final model ended up with an F1 score of 0.91 and an accuracy of 0.94. An small improvement from my base model, but unfortunately it did not improve much classifying "medium" and "low" calorie recipes. Looking back it may have been easier to predict a variable with more variance. Roughly, 58.9% of the average ratings were 5.0 which most likely made prediciting classes outside of "high" difficult.

Fairness Analysis

For my fairness analysis, I split the dataset into low and high calorie recipes. Similar to my hypothesis test, low-calorire recipes are recipes that are below the median and high-calorie recipes are recipes that are above.