View on GitHub

recipe-rating-analysis

This is the EECS 398 Portfolio Homework for Hanlu Wu and Xiulin Chen at the University of Michigan, Ann Arbor. This analysis focuses on predicting the relationships between recipe features and their ratings. The features considered in the analysis include tags, contributor_id, recipe_id, rating, review, minutes, and n_steps.

Data on the Menu: Understanding What Makes a Recipe Shine

Introduction

Dataset Overview

This project analyzes two datasets: a recipe dataset and an interaction dataset, containing information about recipes, user reviews, and ratings. The combined dataset offers rich insights into user preferences, cooking behaviors, and recipe characteristics. Specifically, the dataset contains:

The recipe dataset contains 83,782 rows, and the interaction dataset contains 731,927 rows. These datasets are merged on the recipe_id column to create a unified dataset for analysis.

Research Question

What factors influence a recipe’s average rating, and can we predict it using its characteristics (e.g., cooking time, nutritional values, and tags)?

Understanding the factors that drive user satisfaction with recipes can help chefs and content creators design recipes that better match user preferences. It also provides insight into how various features of a recipe (like preparation time or healthiness) impact its success.

Relevant Columns

Data Cleaning

Overview of Cleaning Steps

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

  1. Merging Datasets:
    • Combined two datasets using “recipe ID” as the key.
    • Calculated the average rating for each recipe to create a new rating_average feature.
  2. Handling Missing Values:
    • Replaced zero values in the rating column with NaN to indicate missing ratings.
    • Removed rows with missing values in critical columns, such as review, tags, and calories, to ensure data completeness for analysis.
  3. Data Type Transformation:
    • Converted the nutrition column (originally a string) into a Python list to allow easy access to individual nutritional components (e.g., calories, fat).
    • Transformed the tags column from a string format into a list of tags for better usability.
  4. Feature Engineering:
    • Created additional features to enhance the dataset:
      • review_length: Measured the word count of each review.
      • num_tags: Counted the number of tags associated with each recipe.
  5. Splitting Nutritional Data:
    • Extracted specific nutrients (e.g., calories, protein, fat, sugar) into separate columns to analyze their influence on recipe ratings.

Cleaned Data display

Relevant Data Sample Head

Below is the head of the cleaned relevant data:

recipe_id tags rating_average review minutes n_steps calories total_fat_pdv sugar_pdv protein_pdv num_tags review_length
333281 ['60-minutes-or-less', 'time-to...'] 4 These were pretty good, but took... 40 10 138.4 10 50 3 14 50
453467 ['60-minutes-or-less', 'cuisine...'] 5 Originally I was gonna cut the rec... 45 12 595.1 46 211 13 9 65

Merged Data Sample

Below is a sample of the merged dataset:

name recipe_id minutes contributor_id submitted tags nutrition n_steps steps description ingredients n_ingredients user_id date rating review rating_average calories total_fat_pdv sugar_pdv protein_pdv
1 brownies in the world best ever 333281 40 985201 2008-10-27 ['60-minutes-or-less', 'time-to-make', ...] [138.4, 10.0, 50.0, ...] 10 ['heat the oven...', 'line...', ...] These are the most chocolatey... ['bittersweet chocolate', 'unsalted ...] 9 386585 2008-11-19 4 These were pretty good, but... 4 138.4 10 50 3
1 in canada chocolate chip cookies 453467 45 1848091 2011-04-11 ['60-minutes-or-less', 'time-to-make', ...] [595.1, 46.0, 211.0, ...] 12 ['pre-heat oven...', 'in a mixing...', ...] This is the recipe we use... ['white sugar', 'brown sugar', ...] 11 424680 2012-01-26 5 Originally I was gonna... 5 595.1 46 211 13

Visualization

Univariate Analysis

The most common tags, such as “preparation,” “time-to-make,” and “course,” are general descriptors commonly used to categorize recipes. These tags likely reflect structural organization rather than specific culinary attributes. Tags like “main-ingredient” and “dietary” are prominent, indicating that recipes are often classified based on their primary components or dietary considerations. This aligns with the importance of catering to specific dietary needs (e.g., vegetarian, gluten-free) in recipe datasets. The tags “easy” and “occasion” highlight that simplicity and context (e.g., festive or special events) are important aspects of recipe selection. These attributes may reflect users’ preferences for quick and contextually appropriate dishes.

Bivariate Analysis

The majority of recipes have high ratings (4-5), indicating they are generally well-received. Low ratings (1-2) are rare, suggesting either fewer poor-quality recipes or under-reporting of negative experiences. The distribution is positively skewed, with a strong bias toward higher ratings. Variability is observed in the mid-range (3-4 ratings), indicating mixed feedback for average recipes.

Across all rating categories, the median preparation time (log-transformed) is relatively consistent, suggesting that preparation time might not strongly influence a recipe’s rating. Lower-rated recipes (categories 0-1 and 1-2) exhibit slightly more variability in preparation time, with some recipes requiring notably longer times. Higher-rated recipes (categories 4-5) show a tighter distribution, implying that recipes with more predictable preparation times tend to receive better ratings. Recipes that are too time-consuming may deter users, especially for lower-rated recipes. Focusing on optimizing preparation time for low-rated recipes could improve their reception.

Specifically , here are what our cleaned data look like among those, relevant data only select the relevant cols in this analysis, and merged_data include all cols we have cleaned up.

relevant_data_head = relevant_data.head().to_markdown(index=False)
merged_data_head = merged_data.head().to_markdown(index=False)

Prediction Problem

Problem Type

Our prediction problem is to forecast the average rating of a recipe based on its features, such as the number of steps, review length, tags, and nutritional content. This is a regression problem because the response variable, rating_average, is continuous and ranges between 0 and 5.

Feature Selection

The following features were used as predictors:

Numerical features were standardized using StandardScaler.

Rationale for Feature Selection:

Response Variable

The response variable is rating_average:

Evaluation Metric

The primary evaluation metric for this regression problem is Root Mean Squared Error (RMSE):

Model Selection and Evaluation

Objective

The goal was to identify the best model for predicting the average rating of recipes (rating_average) using the cleaned and preprocessed dataset. To achieve this, we tested a variety of regression models and evaluated their performance using cross-validation and test set evaluation.

Models Tested

We compared the following models to determine the best-performing one:

  1. Random Forest Regressor: A robust ensemble-based model that reduces overfitting using bagging and feature randomness.
  2. Gradient Boosting Regressor: An ensemble technique that builds models sequentially to correct errors of prior models.
  3. Support Vector Regressor (SVR): A kernel-based regression technique designed for smaller datasets with high complexity.
  4. Ridge Regression: A linear regression model with L2 regularization to prevent overfitting.
  5. Lasso Regression: A linear regression model with L1 regularization for feature selection.
  6. Linear Regression: A simple regression model without regularization.

Methodology

  1. Data Sampling:
    • To speed up computation, we subsampled 20% of the dataset while retaining the overall distribution.
  2. Data Splitting:
    • The data was split into 60% training and 40% test sets using train_test_split.
  3. Cross-Validation:
    • Each model was evaluated using 3-fold cross-validation on the training set.
    • Cross-validation scores were calculated using Root Mean Squared Error (RMSE), which penalizes larger errors more heavily than Mean Absolute Error (MAE).
    • For each model, we computed the mean RMSE and the standard deviation of RMSE across the folds.
  4. Model Selection:
    • The model with the lowest mean RMSE across the cross-validation folds was selected as the best model.

Results of Cross-Validation

Model Mean RMSE ± Std Dev
Linear Regression 0.72 ± 0.02
Random Forest 0.71 ± 0.01
Gradient Boosting 0.72 ± 0.02
Support Vector Regressor (SVR) 0.74 ± 0.02
Ridge Regression 0.72 ± 0.02
Lasso Regression 0.72 ± 0.02

The best-performing model was:

The ** Random Forest ** was selected as the best model for predicting recipe ratings based on its superior performance in cross-validation and test set evaluation. The low RMSE and high R² score indicate that the model captures the variability in the data effectively. This model can now be used for further predictions or integrated into a recommendation system.

Baseline Model: A Foundation to Build On

Objective

The baseline model aims to establish a reference performance using the cleaned dataset. This step serves as a foundation for comparing and improving upon the model in subsequent iterations.

Methodology

  1. Data Splitting:
    • The dataset was divided into 60% training data and 40% test data using train_test_split.
  2. Subsampling:
    • To reduce computation time, 20% of the training data was randomly sampled to train the model. This allows the model to learn efficiently while retaining representativeness.
  3. Model Selection:
    • A RandomForest Regressor was used as the baseline model due to its robustness and ability to handle feature interactions automatically.
  4. Pipeline Construction:
    • A Pipeline was created to combine the preprocessing steps (scaling numerical features) with the model. This ensures consistent preprocessing during training and testing.
  5. Evaluation Metric:
    • The model’s performance was evaluated using Root Mean Squared Error (RMSE), which penalizes larger errors more heavily, making it a suitable metric for regression tasks.

Results

The baseline model provides a benchmark RMSE of 0.74. This indicates that the model has reasonable predictive capability but leaves room for optimization and improvement. Future steps will explore additional features, fine-tune hyperparameters, and test more advanced models to improve performance.

Final Model: Elevating Predictions

Objective

The final model incorporates additional features and hyperparameter tuning to improve performance over the baseline model. The goal is to minimize the prediction error (measured by RMSE) on the test dataset.

Features Added

To enhance the predictive power of the model, the following features were included:

Methodology

1. Data Preprocessing

2. Hyperparameter Tuning

3. Model Training

Results

Metric Baseline Model Final Model
RMSE 0.74 0.70

Insights and Improvements

Conclusion

The final model, a RandomForest Regressor with optimized hyperparameters, demonstrates robust performance in predicting recipe ratings. Future improvements could explore:

About

Authors: