View on GitHub

Recipe Step Predictor

Predicting Recipe Steps: An Analysis of Food.com Recipes

Predicting Recipe Steps: An Analysis of Food.com Recipes is a data science project conducted at University of Michigan. The project encompasses various stages of analysis, starting from exploratory data analysis to creating various models.

Authors: Naman Jain, Anay Moitra

Introduction

General Introduction

The Food.com Recipes and Ratings dataset contains over 80,000 recipes and 700,000 ratings collected from the Food.com website since 2008. Each recipe includes details such as the preparation time, ingredients, steps, nutritional information, tags, and user ratings. This rich dataset provides an excellent opportunity to explore various factors that influence recipe complexity and user preferences.

In the culinary world, understanding what makes a recipe complex or straightforward can significantly impact home cooks, chefs, and recipe developers. The number of steps in a recipe often correlates with the recipe’s complexity, preparation time, and required skill level.

The central question we aim to answer is: What factors influence the number of steps in a recipe?.

By analyzing various features such as ingredients, nutritional content, and tags associated with recipes, we hope to find patterns that can help predict recipe complexity. This analysis could assist home cooks in selecting recipes that match their skill level and available time, and help recipe developers create recipes that cater to their target audience.

Introduction of Columns

The cleaned dataset consists of 83,782 rows of recipes with the following relevant columns:

Data Cleaning and Exploratory Data Analysis

Data Cleaning

To prepare the dataset for analysis, a few data cleaning steps were performed:

Parsing the nutrition Column:

By extracting individual components from the nutrition column, these variables allowed us to include specific nutritional factors as predictors in our models. For example, understanding the relationship between protein_PDV and n_steps helped identify how recipe complexity correlates with nutrition. Without this step, the nutrition data would have been unusable due to its string-based format.

Handling the tags Column:

Converting the tags column into lists and identifying unique tags helped us to engineer categorical features that describe the nature of recipes. For example, one-hot encoding tags like easy and dessert can provide valuable information for predicting n_steps. This step has qualitative aspects of the recipes, which was important in exploring patterns such as whether simpler tags are associated with fewer steps.

Merging Average Ratings:

Handling Missing Values:

Adding average_rating provided an important metric for evaluating the popularity or perceived quality of recipes. Although it had missing values for unrated recipes, leaving these values as NaN avoided introducing bias by imputing artificial ratings. This decision preserved the integrity of the analysis and made sure that if we had done predictions involving average_rating, they would not skewed by incorrect assumptions about missing values.

This is what our cleaned data looks like:

name minutes tags n_steps n_ingredients calories total_fat_PDV protein_PDV
1 brownies in the world best ever 40 [‘60-minutes-or-less’, ‘time-to-make’, ‘course’, ‘main-ingredient’, ‘preparation’, ‘for-large-groups’, ‘desserts’, ‘lunch’, ‘snacks’, ‘cookies-and-brownies’, ‘chocolate’, ‘bar-cookies’, ‘brownies’, ‘number-of-servings’] 10 9 138.4 10 3
1 in canada chocolate chip cookies 45 [‘60-minutes-or-less’, ‘time-to-make’, ‘cuisine’, ‘preparation’, ‘north-american’, ‘for-large-groups’, ‘canadian’, ‘british-columbian’, ‘number-of-servings’] 12 11 595.1 46 13
412 broccoli casserole 40 [‘60-minutes-or-less’, ‘time-to-make’, ‘course’, ‘main-ingredient’, ‘preparation’, ‘side-dishes’, ‘vegetables’, ‘easy’, ‘beginner-cook’, ‘broccoli’] 6 9 194.8 20 22
millionaire pound cake 120 [‘time-to-make’, ‘course’, ‘cuisine’, ‘preparation’, ‘occasion’, ‘north-american’, ‘desserts’, ‘american’, ‘southern-united-states’, ‘dinner-party’, ‘holiday-event’, ‘cakes’, ‘dietary’, ‘christmas’, ‘thanksgiving’, ‘low-sodium’, ‘low-in-something’, ‘taste-mood’, ‘sweet’, ‘4-hours-or-less’] 7 7 878.3 63 20
2000 meatloaf 90 [‘time-to-make’, ‘course’, ‘main-ingredient’, ‘preparation’, ‘main-dish’, ‘potatoes’, ‘vegetables’, ‘4-hours-or-less’, ‘meatloaf’, ‘simply-potatoes2’] 17 13 267 30 29

Univariate Analysis

We start with performing univariate analysis to examine the distribution of single variables. We looked at the distribution of steps in recipes which is show below:

The plot shows that most recipes have a relatively low number of steps, with a sharp decline as the number of steps increases. This suggests that simpler recipes (with fewer steps) are far more common, and complex recipes with many steps are rare. This observation also highlights the prevalence of simple recipes and indicates that factors like ingredient count may play a significant role in determining the complexity of a recipe.

Bivariate Analysis

We then performed bivariate analysis to examine the relationship of different variables. We looked at the relationship between the number of steps and the number of ingredients in recipes which is show below:

The trend shows that as the number of ingredients increases, the number of steps tends to spread out, with some recipes having significantly more steps. However, the relationship is not strictly linear, which suggests that factors other than ingredient count, such as preparation complexity or cooking techniques, may also influence the number of steps in a recipe. This supports our investigation into what drives recipe complexity.

Interesting Aggregates

We also created a pivot table to examine the relationship between the number of ingredients and the average number of steps in recipes. This analysis helps identify how the recipe complexity changes with varying ingredient counts. The first few rows of the pivot table are shown below:

n_ingredients n_steps
1 7.57143
2 5.93173
3 5.61315
4 6.32002
5 7.12584

This pivot table shows that recipes with fewer ingredients tend to have fewer steps on average, but the relationship is not strictly linear. For example, recipes with one ingredient have a higher average number of steps than those with two or three ingredients. This could indicate that recipes with very few ingredients might involve more complex cooking techniques, while those with moderate ingredient counts might be simpler to prepare.

Imputation

We also saw that some of our columns had missing values. These columns were name, description, and average rating. We decided to not impute these values as none of these columns were relavent to us and imputing or not imputing values would have no effect on our analysis.

Framing a Prediction Problem

With a solid understanding of our dataset and initial exploratory analyses, we now defined a clear prediction problem that aligns with our overall theme of understanding recipe complexity.

Prediction Problem Statement

We aim to predict the number of steps in a recipe (n_steps) using features that are available before the recipe steps are fully finalized. By estimating the complexity of a recipe from its attributes, we can help home cooks choose recipes appropriate for their time constraints and skill levels, and assist recipe creators in refining their dishes for targeted audiences. Since our response variable, n_steps, is a continuous numerical value, this is a regression problem, since we are predicting a numerical outcome.

We chose n_steps because it serves as a direct measure of recipe complexity. Understanding what influences complexity can reveal insights into how preparation time, nutritional factors, and the diversity of tags (dietary restrictions, meal types, cuisines) shape the effort required to prepare a dish.

Features and Time of Prediction Justification

Here are some of the features we will use to predict n_steps:

By relying solely on information available prior to detailing the step-by-step instructions, we made that our prediction does not leak information from the future. Basically, we are not using features that depend on knowing the final number of steps or user feedback that would only be available after the recipe is published.

Evaluation Metric

We will use both Mean Absolute Error and R-squared (R²) as our evaluation metric.

R² is a standard metric for regression tasks and has several advantages. One of them is R² indicates the proportion of the variance in the target variable that is explained by the model, which provides a clear understanding of the model’s explanatory power. R² values range from 0 to 1 (or can be negative for poor models), which also makes it easy to interpret how well the model fits the data. A higher R² means the model’s predictions are more aligned with the actual values. MAE is a metric used to measure the average magnitude of errors between predicted values and actual values and also has many advantages in this scenario. One of them is MAE is easy to understand because it’s the average of absolute differences between predicted and actual values. This gives a direct, intuitive sense of how far off the model’s predictions are, on average.

Together, R² will help us understand how well the model captures the overall variance in the data, while MAE will provide a sense of the average error magnitude, which will make sure both model fit and prediction accuracy are considered.

With our prediction problem defined, our response variable chosen, and our metric justified, we have a clear path forward. Next, we will build baseline and final models to predict n_steps and evaluate how well our model performs in capturing the complexity of recipes.

Baseline Model

Our baseline model is a linear regression model aimed at predicting the number of steps (n_steps). We used two quantitative features: total_fat_PDV (quantitative) and protein_PDV (quantitative). These features were chosen based on the assumption that recipes with higher fat and protein content might be more complex and require a greater number of steps.

After fitting the model, the R² value on the test set was 0.027, which indicates that only 2.7% of the variance in n_steps was explained by these features. Although this is quite low, the MAE was 4.5363, which suggests the model’s predictions were off by an average of about 4.5 steps. This result shows that the selected features provide only a weak relationship with recipe complexity but still offer a rough approximation of the number of steps.

While the baseline model’s performance leaves a lot of room for improvement, it establishes a foundation for development. For our final model, we will explore adding more features and applying hyperparameter tuning to enhance predictive performance and better capture the complexity of recipes.

Final Model

In our final model, we added two more features: n_ingredients (quantitative) and calories (quantitative). We believe these two features have a high impact on the number of steps as when increasing the number of ingredients of a dish, we expect utilizing all those ingredients would typically increase the number of steps. We also added calories for a similar reason as using total fat and protein.

The modeling process involved testing multiple algorithms, including Ridge Regression, Lasso Regression, ElasticNet, and Gradient Boosting Regressor. These models were selected to balance flexibility and regularization. The hyperparameters for each model were tuned using GridSearchCV with a robust cross-validation strategy (RepeatedKFold) to make sure that the model performed well on unseen data. For ElasticNet, the best hyperparameters were an alpha value of 0.1 for regularization strength and an l1_ratio of 0.5, balancing L1 and L2 penalties. For the Gradient Boosting Regressor, the best hyperparameters included a learning_rate of 0.1, 200 n_estimators, and a max_depth of 5, which allowed the model to capture some of the complex, nonlinear relationships in this dataset.

Our new Mean Absolute Error is down to 3.984 showing our estimations are only about 4 steps off on average. Our R² value is now 0.2192, meaning our model has improved quite a lot in this aspect as well. We have achieved some improvement in both of our evaluation metrics, suggesting that we made an effective adjustment in our final model. These metrics show that the final model provides a more accurate and reliable prediction of n_steps. We believe that this improvement occured because the number of ingredients in a recipe inherently relates to its complexity and the calories per serving reflect the overall energy density of a dish, which often corresponds to the richness and variety of ingredients, thus this feature provides an indirect measure of recipe complexity that complements n_ingredients.

We also experimented with one-hot encoding the tags column, but it did not significantly improve the final model’s performance, which is likely due to the high complexity and sparsity of the tags. This may not have effectively captured meaningful patterns related to n_steps.

To conclude and to answer our question of what factors influence the number of steps in a recipe, our analysis shows that the number of ingredients and nutritional factors like total_fat_PDV, protein_PDV, and calories play a very important role in predicting recipe complexity. Recipes with higher ingredient counts and greater nutritional values tend to involve more steps, though the relationship is not strictly linear.

While our final model showed improved performance over the baseline, the relatively modest R² tells us that there are still unexplored factors influencing recipe complexity. Exploring tags and additional data columns might give us further insights into this, and more advanced models could potentially catch the nuances in this dataset. Nevertheless, our findings show the importance of nutritional and ingredient-related features in understanding recipe complexity.