---
dataset_info:
  pretty_name: Recipe Cleaned Dataset
  task_categories:
    - recommendation
    - text-classification
  languages:
    - en
  size_categories:
    - 100K<n<1M
  license: mit
  annotations_creators:
    - machine
  source_datasets:
    - original
  multilinguality: monolingual
---

# Recipe Cleaned Dataset

## Dataset Summary
This dataset is a structured and cleaned collection of recipe data derived from the [Food.com Recipes and Interactions dataset](https://www.kaggle.com/datasets/shuyangli94/food-com-recipes-and-user-interactions/data?select=RAW_recipes.csv). It is designed for ingredient-based personalization, machine learning training, and interactive recommendation systems. The dataset integrates a hierarchical ingredient taxonomy, standardized nutrition information, and categorical metadata (e.g., diet tags, cuisine attributes, region) to support downstream filtering and modeling.

---

## Source and Collection Process

The original data comes from Food.com via Kaggle. It consists of recipes and user interactions collected over several years, published under a permissive license for research and educational use. Only the `RAW_recipes.csv` portion was used in this dataset. No private or login-restricted data was collected.

---

## Labeling and Cleaning Process

All labeling was performed using a **rule-based and embedding-based taxonomy mapping pipeline**. Ingredients were clustered into parent classes using sentence embeddings (~1,000+ parent nodes), and child phrases were mapped using keyword expansion and string matching. A small subset was manually inspected to ensure mapping accuracy.

Additional processing includes:
- Deduplication and invalid entry filtering
- Text normalization
- Region and diet tag standardization
- Nutrition field standardization

---

## Ethical Notes

All data are derived from a publicly released Kaggle dataset.  
No personal user information is included.  
Nutrition values are estimated and must **not** be used for health or medical decisions.

---

## Augmented/Synthetic Data

No augmented or synthetic samples are included in this dataset.

---

## Dataset Structure

- **Format**: CSV (UTF-8)  
- **Number of Recipes**: ~180,000  
- **Schema**:
  - **Basic Info**: `name`, `minutes`, `tags`, `n_steps`, `steps`
  - **Ingredients (raw & standardized)**:
    - Raw ingredient lists split into categories: `staple`, `main`, `seasoning`, `other`
    - Each category has a mapped parent version: e.g. `staple_parent`, `main_parent`, etc.
    - `n_ingredients` gives the count of ingredients per recipe.
  - **Nutrition**: `calories`, `total_fat_pdv`, `sugar_pdv`, `sodium_pdv`, `protein_pdv`, `saturated_fat_pdv`, `carbohydrates_pdv`  
    (all standardized per serving; PDV = % Daily Value)
  - **Metadata**: `region` (51 classes), `diet_tags`, `cuisine_attr` (671 classes), `is_vegan_safe`, `is_vegetarian_safe`

---

## Ingredient Taxonomy

To standardize heterogeneous ingredient strings, a parent–child taxonomy was constructed:

- Approximately 1,000+ parent ingredient classes generated through sentence embedding clustering.
- Original phrases (child nodes) are mapped via rule-based string matching and keyword expansion.
- Four functional categories are defined:
  - **Staple**: e.g., rice, pasta, flour
  - **Main**: protein sources or key components
  - **Seasoning**: spices, condiments, sauces
  - **Other**: toppings, garnishes, optional extras

This structure enables consistent representation across recipes, improves filtering granularity, and simplifies feature engineering for ML models.

---

## Data Cleaning Pipeline

1. Deduplication & Filtering — remove exact/near-duplicates and invalid rows.  
2. Text Normalization — lowercase conversion, punctuation cleanup, Unicode fixes.  
3. Ingredient Categorization & Mapping — split into functional categories, map to parent classes.  
4. Tag & Region Standardization — unify tags and infer cuisine/region attributes.  
5. Nutrition Standardization — align units, compute PDV values, fill missing fields when possible.

---

## Risks and Bias

The dataset inherits several biases from the original Food.com source:

- **Regional bias**: Western and American cuisines dominate, leading to under-representation of other cuisines.
- **Popularity bias**: Highly rated and frequently interacted recipes are over-represented, which may skew recommendation models.
- **Nutritional estimation bias**: Nutrition values are approximate and may vary from real-world measurements.

These biases may affect fairness and diversity in downstream models.

---

## Intended Uses and Limitations

**Intended Uses**
- Ingredient-based recipe recommendation research
- Cold-start personalization and ranking model development
- Educational and academic applications

**Not Intended For**
- Medical or nutritional decision-making
- High-stakes personalization without bias mitigation
- Commercial deployment without further validation

---

## Exploratory Data Analysis (EDA)

### Sample Entry

```json
{
  "name": "Spaghetti Aglio e Olio",
  "minutes": 15,
  "ingredients": ["spaghetti", "garlic", "olive oil", "red pepper flakes", "parsley", "salt"],
  "main_parent": ["pasta"],
  "seasoning_parent": ["garlic", "olive oil", "red pepper", "salt"],
  "calories": 420,
  "region": "Italian",
  "diet_tags": ["vegetarian"],
  "n_steps": 3
}