View on GitHub

recipes-and-cuisines-analysis

Analysis of cuisines and other factors from a large recipes dataset

Analyzing Recipes and Cuisines

Analysis of cuisines and other factors from a large recipes dataset

By Rhys Burman and Kavya Narayan

Introduction

Cooking is not just a necessity but a universal language that bridges cultures and traditions. From haphazardly learning to make mac n’ cheese as an underfunded college student to making large Thanksgiving dinners for the whole family, the art of cooking has always bound us all together. In the age of globalization, understanding the relationships between cuisines and recipe components (e.g., ingredients, preparation time, health considerations) can reveal fascinating insights into how food connects us all. This project explores a dataset of recipes to answer one central question:

“Can we predict the cuisine of a dish based on its ingredients and preparation details?”

Understanding this question has implications for multiple audiences:

Dataset Overview

The datasets we used contain a subset of recipes and user reviews collected from Food.com, including only recipes and reviews posted since 2008, given that the original dataset was quite large. These datasets offered a comprehensive view of recipe characteristics and user preferences, making it ideal for our exploration.

Key statistics about the datasets:

In our exploration, we merged both datasets to create one dataset with the recipes and their respective reviews.

Relevant columns from each of the datasets:

For the Recipes Dataset:

Column Description
name Recipe name
id Recipe ID
minutes Minutes to prepare recipe
tags Food.com tags for recipe
nutrition Nutrition information in the form [calories (#), total fat (PDV), sugar (PDV), sodium (PDV),
  protein (PDV), saturated fat (PDV), carbohydrates (PDV)]; PDV stands for “percentage of daily
  value”
steps Text for recipe steps, in order
description User-provided description
ingredients The ingredients used in the recipe
n_ingredients The total number of ingredients used in the recipe

For the Reviews/Ratings Dataset:

Column Description
user_id User ID
recipe_id Recipe ID
rating Rating given
review Review text

Why Should You Care?

The dataset and the central question are highly relevant for anyone interested in food, technology, or cultural exploration. Imagine a system that can recommend dishes from diverse cuisines based on the ingredients in your fridge! Beyond individual convenience, we hope our analyis contributes to cross-cultural appreciation and dietary innovation.

Stay tuned as we explore this dataset and uncover the secrets of global cuisines through data analysis and machine learning!

Data Cleaning and Exploratory Data Analysis

Data Cleaning:

  1. In order to prepare the data for our analysis, we first left merged the recipes and reviews on the recipes dataset to create one complete dataset, named merged.
  2. We then filled all ratings of 0 with np.Nan because there is no option to rate a recipe as a 0, so a 0 value for the rating implies that the reviewer did not leave a rating for this recipe. Therefore, it makes more sense to change it to np.Nan so that the 0 values did not skew our later calculations for the average ratings for the recipe.
  3. Next, we dropped all of the columns we did not believe were relevant for our analysis. These included contributor_id submitted, n_steps, n_ingredients, user_id, recipe_id, date, review. Since we were focusing on cuisines in particular, we did not think any of these columns were relevant.
  4. After removing irrelevant columns, we calculated the average rating for each recipe. To do this, we grouped the dataset by the recipe ID (id) and took the mean of the ratings for each recipe. We then added this average rating back to the dataset as a new column, average_rating.
  5. The nutrition column in the dataset was originally stored as a string representation of a list containing various nutritional values (e.g., [calories, total fat, sugar, sodium, protein, saturated fat, carbohydrates]). To make these values usable for analysis, we converted this string into an actual list using ast.literal_eval and then split it into individual columns for each nutritional component. This created new columns such as calories, total_fat, sugar, sodium, protein, saturated_fat, and carbohydrates.
  6. We then calculated a custom health_rating for each recipe. The health rating was computed based on the nutritional values of the recipes, with higher calories, saturated fat, sugar, sodium, and carbohydrates reducing the score, while higher protein content increased the score. In our equation, we divided each of these nutritional values by the mean of the nutritional value across all of the recipes in the dataset (for e.g: calories was divided by the mean calories for all recipes which was added to total fat divded by the mean of all total fat values) This normalization ensured that recipes with healthier nutritional profiles received higher scores. In calculating these, we gave each dish a score of 10 to begin with and then subtracted each of these values from the initial score.
  7. The tags column, which contained various descriptive tags about each recipe, was initially stored as a string representation of a list. We converted it into an actual list format using ast.literal_eval to enable easier analysis. We then cross-referenced the tags with a predefined list of cuisines stored in a separate CSV file. Using this, we created a new column, cuisines, which contained only the relevant cuisine tags for each recipe.
  8. Then, we dropped columns that were no longer needed for our analysis, such as the original nutrition column, individual nutritional components (calories, total_fat, etc.), and tags. Additionally, we removed rows where the cuisines column was empty or missing values to ensure that our dataset only included relevant and complete data for analysis. Ultimately, this left us with 8 columns and 72,240 rows.
  9. Finally, we groupby for name so we had unique recipes with their associated cuisines in our dataset, leaving us with 25,797 rows, or in other words, 25,797 unique recipes.

By the end of this cleaning process, the dataset was reduced to include only the essential columns required for our analysis: minutes, description, steps, ingredients, average_rating, health_rating, and cuisines. This cleaned dataset provided a strong foundation for conducting exploratory data analysis and building predictive models. The head of our final cleaned merged dataframe is shown below:

(For better readability, the outputs are truncated):

minutes steps description ingredients average_rating health_rating cuisines
5 [‘rinse martini glass with extra dry vermouth’, ‘a…] based on a recipe from ray foley’s the ultimate li… [‘dry vermouth’, ‘vodka’, ‘gin’, ‘lillet blanc’] 5 9.64705 european
25 [‘combine cinnamon , vanilla , cornstarch , peach …] heart healthy [‘cinnamon’, ‘vanilla extract’, ‘cornstarch’, ‘pea…] 4 6.31694 african
15 [‘preheat oven to 350 degrees , and line a baking …] i came across these cookies once by accident, and … [‘natural-style peanut butter’, ‘sugar substitute’…] 2 6.71192 canadian
50 [‘mix batter ingredients in a 9-1 / 2-inch deep di…] a quick recipe with only 20 minutes preparation us… [‘cooking spray’, ‘all-purpose flour’, ‘fast risin…] 5 7.94085 european, american, italian
45 [‘pre-heat oven the 350 degrees f’, ‘in a mixing b…] this is the recipe that we use at my school cafete… [‘white sugar’, ‘brown sugar’, ‘salt’, ‘margarine’…] 5 1.53134 canadian

Univariate Analysis:

First, we wanted to look at the distribution of cuisines, as is seen below.

From this graph, we could tell there was an imbalance in the dataset, with American and European cuisines heavily dominating (around 8000 entries each) while many other cuisines like Korean, Malaysian and Thai have fewer than 100 entries each. This imbalanced distribution suggested to us that while predicting cuisine from ingredients might be possible, our model would likely show bias towards American and European cuisines so we needed to ensure we specifically address the class imbalance issue.

Next, since our model will be predicting the cuisine at least partially based on the ingredients, we wanted to see the distribution of the most common ingredients as well.

Looking at the distribution of ingredients across recipes, we found that salt is overwhelmingly the most common ingredient with over 10,000 appearances, followed by a cluster of fundamental ingredients like olive oil, butter, onion, and garlic cloves each appearing in around 5,000 recipes. This pattern suggested to us that these basic ingredients form the foundation of cooking across many cuisines, which could make them less useful as distinguishing features for predicting specific cuisines, and might be something to look out for in creating our model.

Bivariate Analysis:

Next, we conducted some bivariate analyses in order to examine relationships between cuisines and other variables in the dataset. First, we looked at if different cuisines had different health ratings, as this was a variable we was planning on training our model with. The graph of the relationship between cuisine and health rating is shown below:

Examining the average health ratings across cuisines, we found that Pakistani and Turkish cuisines rank highest with scores above 7, while Irish cuisine ranks lowest at 4.19. Interestingly, certain Asian cuisines like Korean, and Chinese fall in the middle range around 6-7. From our definition and calculation of health rating, higher health ratings indicate healthier dishes, which indicated to us that some cuisines tended to have healthier dishes than others. Although this distribution is relatively balanced, there are differences between these cuisines that did prove useful in establishing our model.

Next, we also wanted to look at preparation times for each cuisine - the graph is shown below:

From this plot, we can see German and Korean cuisines stand out as requiring significantly more preparation time, averaging around 270-280 minutes, while Filipino and Australian dishes are the quickest to prepare at under 50 minutes. This substantial time variation between cuisines (from 50 to 280 minutes) was an important feature for cuisine prediction, because it reflected fundamental differences in cooking techniques and complexity.

Finally, since we predicted that different cuisines would use different ingredients, we also wanted to make some graphs to see the most common ingredients for different cuisines.

The heatmap and scatter plot revealed distinct ingredient usage patterns across different cuisines. European and American cuisines show the highest diversity and frequency of ingredient usage (as shown by the more frequent and brighter spots in the heatmap and higher dots in the scatter plot), while cuisines like Malaysian, Korean, and Turkish have more concentrated and specific ingredient profiles (shown by fewer but distinct ingredient patterns). These trends supported our hypothesis that ingredients can be useful predictors of cuisine type, though the strong overlap in common ingredients (like salt, butter, and onions) across multiple cuisines suggested to us that we’d need to rely on the unique combinations of ingredients rather than individual ingredients alone for accurate cuisine prediction.

Interesting Aggregates:

To better understand the relationship between cuisines, health ratings, preparation times and average star ratings, we created a pivot table that shows the average health rating and preparation time for each cuisine type. The table reveals some interesting patterns:

cuisines Average Health Rating Average Preparation Time (mins)
pakistani 7.32 58.26
turkish 7.26 49.52
indian 7.01 70.18
spanish 6.9 125.65
central-american 6.87 91.21
middle-eastern 6.85 113.93
greek 6.77 66.79
chinese 6.75 73.11
russian 6.7 78.79
mexican 6.64 101.2
italian 6.48 71.85
korean 6.34 292.76
african 6.23 113.18
french 6.22 208.66
australian 5.97 48.12
portuguese 5.94 187.67
european 5.9 104.28
german 5.89 285.92
scandinavian 5.78 73.47
canadian 5.75 134.86
american 5.69 92.77
filipino 5.61 48.67
malaysian 5.52 50.93
hawaiian 5.35 99.92
vietnamese 5.11 70.65
thai 4.93 65.85
southern-united-states 4.82 121.26
caribbean 4.76 98.74
japanese 4.66 70.07
south-african 4.53 78.73
irish 4.08 202.21

This aggregate analysis revealed several interesting insights:

  1. South Asian cuisines (Pakistani, Indian) tend to have higher health ratings on average
  2. Korean cuisine stands out with the longest average preparation time (292.76 minutes)
  3. Australian cuisine combines a moderate health rating with relatively quick preparation times
  4. There’s no strong correlation between preparation time and health rating, suggesting these are independent characteristics of different cuisines
  5. Our results for health ratings and preparation times are slightly different for this table in comparison to our initial bivariate graphs, likely because in for this table, we used a more nuanced method in dropping cuisines when they had more than one associated cuisine. This time, instead of simply dropping the second cuisine, we instead dropped the cuisine that was more popular of the two associated cuisines for each dish, so that minority cuisines had greater representation.

Overall, these patterns helped inform our model so we would know when adding numerical features into it, which would have a strong impact and which would have less of one as features in our classification model.

Imputation:

In our analysis we did not impute any values as all of our columns of interest already had all of the data we needed and had already been cleaned, and thus we felt as though we could just move on with our prediction.

Framing a Prediction Problem

This is a multiclass classification problem where we aimed to predict the cuisine category of a dish (e.g., American, European, Asian, etc.) using features that would be known before the cuisine is identified. Our response variable was the cuisines column, which we chose because it directly addresses our research question of whether we can identify a dish’s culinary origin based on its characteristics. For model evaluation, we used the F1-score rather than simple accuracy because:

  1. Our dataset was heavily imbalanced (as shown in our univariate analysis where American/European cuisines dominate)
  2. F1-score provides a balanced measure of precision and recall, which is crucial when dealing with imbalanced classes
  3. It would help ensure our model performs well across all cuisine types, not just the dominant ones

The features we used for prediction included:

We specifically excluded features like user ratings and reviews since these would only be available after a recipe is already categorized with its cuisine. This ensured our model only used information that would be available at the “time of prediction” - when someone is trying to identify a cuisine based on a new recipe’s characteristics.

Baseline Model

Our model used a Random Forest Classifier with features that included both quantitative and categorical variables:

To handle these different types of features, we implemented a preprocessing pipeline that:

  1. Applies StandardScaler to the ‘minutes’ feature to normalize the numerical data
  2. Uses TF-IDF vectorization for the ‘ingredients’ feature, converting the text data into a numerical matrix with 5000 features
  3. Addresses class imbalance using balanced class weights, which we computed based on the training data distribution

The model achieved an accuracy of 46.3% on the test set, with varying performance across different cuisines. Looking at the classification report:

We would not consider this current model “good” because:

  1. The overall accuracy of 45.1% is relatively low for practical applications
  2. The model completely fails to predict several cuisines
  3. There’s significant variation in performance across different cuisines, indicating bias towards better-represented classes

Final Model

Based on our analysis of the recipe data and understanding of cuisine classification, we significantly enhanced our Final Model by incorporating several key features and improvements:

Features Added and Their Significance:

  1. Text-Based Features:
    • Recipe Names: We included recipe names because they often contain strong cultural and regional indicators (e.g., “Pad Thai” or “Beef Bourguignon”) that directly hint at their cuisine origins.
    • Recipe Descriptions: Descriptions frequently contain cultural context, traditional serving suggestions, and historic background that can help identify a cuisine’s origin.
    • Cooking Steps: We added detailed cooking steps because cooking techniques vary significantly across cuisines (e.g., stir-frying in Chinese cuisine vs. slow-cooking in French cuisine).
  2. Additional Numerical Features:
    • Health Rating: Based on our bivariate analysis showing distinct nutritional patterns across cuisines (e.g., Pakistani cuisine having higher health ratings than Irish cuisine), we included this as a discriminating feature.
  3. Enhanced Ingredients Processing:
    • We developed a custom IngredientsVectorizer that better captures the nuanced relationships between ingredients and cuisines.
    • Limited to top 1000 ingredients to focus on the most significant ingredients while reducing noise from rare or misspelled items.
    • Used binary encoding to capture ingredient presence, as quantities are less relevant for cuisine classification.

Model Architecture and Hyperparameter Selection:

We chose a Random Forest Classifier as our base algorithm due to its ability to:

Key hyperparameter choices included:

Performance Improvements:

Our Final Model showed substantial improvements over the Baseline Model:

While some very rare cuisines (e.g., Filipino, central-American) still show poor performance due to limited training data, our model’s overall weighted average F1-score improved from 0.40 to 0.54, indicating more balanced and robust performance across cuisines. This improvement suggests that our comprehensive feature set better captures the complex relationships between recipe characteristics and cuisine classifications.

The success of these enhancements aligns with our understanding of how cuisines are differentiated in the real world - not just by ingredients alone, but by the combination of ingredients, cooking methods, cultural context, and nutritional profiles. Our Final Model better reflects this multifaceted nature of cuisine classification.