Food Recipe Insights

Data analysis and prediction model for food recipe ratings. Final project for EECS 398 at University of Michigan.

Project Overview

We’re analyzing the Food.com recipe and rating dataset to understand what factors influence recipe ratings. The dataset contains 83,782 recipes with user ratings and detailed characteristics about each recipe.

Our analysis focuses on a key question: What factors influence a recipe’s rating on Food.com? This question is particularly relevant in today’s digital age, where home cooks increasingly rely on online recipes and reviews to guide their culinary adventures.

Dataset Overview

Number of recipes: 83,782
Key features we’ll examine:
- Average rating (our target variable)
- Cooking time
- Number of steps and ingredients
- Nutritional information
- Recipe categories

Relevant Columns

Our analysis focuses on these key features:

avg_rating: The average user rating for each recipe (1-5 stars)
minutes: Total cooking time required
n_steps: Number of preparation steps
n_ingredients: Number of ingredients used
calories: Caloric content per serving
protein_pdv: Protein content as percentage of daily value
total_fat_pdv: Total fat content as percentage of daily value
carbohydrates_pdv: Carbohydrate content as percentage of daily value
review_count: Number of reviews received
difficulty_score: Calculated score indicating recipe complexity

Data Cleaning and Exploratory Data Analysis

Dataset Overview

Recipe Complexity Distribution

The distribution of recipe complexity shows that most recipes fall into the medium difficulty range (score between 0.7-1.3). There’s a slight positive skew, indicating that Food.com tends to favor more approachable recipes. This aligns with the platform’s focus on home cooking, where overly complex recipes might deter users.

Cooking Time Impact

Analysis of cooking time versus ratings reveals an interesting pattern: recipes with medium cooking times (30-60 minutes) tend to receive higher ratings. Very quick recipes (<15 minutes) and very long recipes (>120 minutes) show slightly lower average ratings, suggesting users prefer recipes that balance convenience with proper cooking time.

Relationship Analysis

Health Metrics Correlation

The relationship between nutritional values shows strong correlations between certain health metrics. Total fat and saturated fat show a strong positive correlation (0.85), while protein and carbohydrates show a weaker relationship (0.32). This suggests that recipes tend to cluster into distinct nutritional profiles.

Key Findings from EDA

Distribution of cooking times
Relationship between ingredients and ratings
Nutritional content patterns

Interesting Aggregates

Recipe Ratings by Cooking Time

Cooking Time Category	Average Rating	Recipe Count	Average Minutes
Very Quick (<15m)	4.67	16,303	9.33
Quick (15-30m)	4.62	20,115	24.87
Medium (30-60m)	4.61	24,570	45.68
Long (60-120m)	4.63	11,840	81.71
Very Long (>120m)	4.59	8,344	779.34

Recipe Complexity Analysis

Cooking Time Category	Avg Ingredients	Avg Steps	Avg Calories
Very Quick (<15m)	6.48	5.55	313.49
Quick (15-30m)	8.88	9.34	375.64
Medium (30-60m)	10.09	11.49	445.81
Long (60-120m)	10.97	13.14	558.00
Very Long (>120m)	10.19	12.31	553.61

Missing Value Analysis

Column	Missing Count	Missing Percentage
name	1	0.00%
description	70	0.08%
avg_rating	2,609	3.11%
cooking_time_category	1	0.00%
time_category	1	0.00%

Note: Only columns with missing values are shown. All other columns have complete data.

Key insights from this data:

Very quick recipes (<15 minutes) have the highest average rating (4.67)
Longer cooking times correlate with more complex recipes (more ingredients and steps)
Missing data is minimal, with only 3.11% missing ratings being the highest percentage
Recipe complexity generally increases with cooking time, but plateaus for very long recipes

Data Imputation

We chose not to impute missing values in our analysis because:

The percentage of missing values was minimal (<2% across key columns)
Missing values appeared to be Missing Completely at Random (MCAR)
For our analysis questions, removing recipes with missing values wouldn’t introduce significant bias
Maintaining data authenticity was crucial for accurate rating prediction

Prediction Problem

Problem Definition

We aim to predict recipe ratings on a scale of 1-5 based on recipe characteristics that are available when a recipe is first posted. This is framed as a regression problem because:

The target variable (avg_rating) is continuous on a 1-5 scale
We need to predict exact rating values rather than categories
The ratings represent a meaningful numeric scale of user satisfaction

Model Components

Target Variable: avg_rating
- Scale: 1-5 stars
- Represents average user satisfaction with recipe
- Dataset size: 83,782 recipes with ratings
Selected Features
- Recipe Characteristics:
  - n_ingredients: Number of ingredients
  - minutes: Cooking time
  - n_steps: Number of preparation steps
- Nutritional Information:
  - calories: Total calories
  - protein_pdv: Protein (% daily value)
  - total_fat_pdv: Fat (% daily value)
  - sodium_pdv: Sodium (% daily value)
- Categorical Features:
  - cooking_time_category: Duration category
Feature Selection Rationale We carefully selected features that would be available at the “time of prediction” (when a recipe is first posted). Specifically excluded:
- review_count: Not available for new recipes
- User-generated tags: May not be present initially
- Submission date: To avoid temporal bias
Evaluation Metric: Mean Squared Error (MSE)
- Chosen because it:
  - Penalizes larger prediction errors more heavily
  - Provides interpretable results in the same units as ratings
  - Aligns with our goal of accurate numerical predictions

Baseline Model

Model Description

We developed a Random Forest Regression model to predict recipe ratings. Our model processes both quantitative and nominal features through a preprocessing pipeline that includes scaling and encoding steps.

Feature Composition

Quantitative Features (2):
- Number of ingredients (n_ingredients)
- Cooking time in minutes (minutes)
Nominal Features (1):
- Cooking time category (cooking_time_category)
  - Encoded categories: very_quick, quick, medium, very_long

Data Processing

Numeric Features:
- Applied median imputation for missing values
- Standardized using StandardScaler
Categorical Features:
- Used OneHotEncoder with drop=’first’
- Imputed missing values with ‘medium’ category

Model Performance

Our Random Forest Regressor achieved the following metrics:

Test RMSE: 0.6482 (average prediction error in rating points)
Train RMSE: 0.6263
Test R²: -0.0392
Train R²: 0.0481

Feature Importance Analysis

Cooking time (minutes): 58.25%
Number of ingredients: 40.71%
Cooking time categories: ~1% combined

Model Assessment

The current model’s performance indicates room for improvement:

The negative R² score suggests the model performs worse than a horizontal line
Similar train and test RMSE values indicate consistent but suboptimal performance
Feature importance analysis reveals that temporal features (cooking time) have the strongest predictive power

This baseline model provides valuable insights but needs enhancement through:

Additional feature engineering
Incorporation of nutritional information
Exploration of interaction terms
Advanced algorithm tuning

Final Model

Feature Engineering

We introduced several carefully engineered features designed to capture key aspects of recipes that could influence ratings:

Recipe Complexity Metrics
- steps_per_ingredient: Measures recipe intricacy by relating number of steps to ingredients
- time_per_step: Indicates how much time each step takes on average
- time_per_ingredient: Shows time investment per ingredient

Rationale: Complex recipes with efficient time usage may be better documented and tested, potentially leading to higher user satisfaction.

Nutritional Balance Features
- protein_fat_ratio: Relationship between protein and fat content
- health_score: Weighted combination of nutritional elements (protein, saturated fat, sugar, sodium)

Rationale: Recipes with balanced nutritional profiles may better meet user expectations for both taste and health.

Model Selection and Optimization

Algorithm: Random Forest Regressor

Chosen for its ability to:
Capture non-linear relationships between features
Handle both numerical and categorical inputs
Provide feature importance rankings
Resist overfitting through ensemble learning

Hyperparameter Optimization:

Used GridSearchCV with 3-fold cross-validation
Best parameters found:
n_estimators: 200
max_depth: 10
min_samples_split: 10
min_samples_leaf: 4

Feature Importance

Top predictive features:

health_score (23.1%)
protein_fat_ratio (21.1%)
time_per_step (14.0%)
steps_per_ingredient (13.2%)
time_per_ingredient (13.2%)

Performance Improvement

The final model showed improvements over the baseline:

Test RMSE: 0.6482 → 0.6360 (1.9% improvement)
Test R²: -0.0392 → -0.0003 (99.2% improvement)

The improvements, while modest in RMSE, show significant gain in R² score, suggesting our engineered features better capture the factors that influence recipe ratings. The dominance of nutritional and complexity metrics in feature importance rankings validates our feature engineering approach.

Note: All features used are available at recipe creation time, ensuring our model can make predictions for new recipes before they receive any ratings.