🍲 Recipe Rescue

What actually makes a recipe get 1-star reviews — complexity, category, or something else?

This project uses the public Food.com recipes-and-interactions dataset from Kaggle (recipe details in RAW_recipes.csv, user ratings and reviews in RAW_interactions.csv) to analyze whether complexity (ingredient count, step count, total time) predicts rating, and whether that relationship differs by category.

Built as a hands-on project to practice: Python fundamentals, NumPy vectorized operations, Pandas DataFrames, boolean/filtering logic, joining and aggregating data, and reshaping/pivoting data — plus an optional small-scale web-scraping exercise.

---

🎯 Project Questions

• Do recipes with more ingredients/steps rate worse, better, or about the same?
• Does the complexity-rating relationship differ across categories (desserts vs. quick dinners)?
• Are there "hidden gem" recipes — high rating, low review count?
• Is total time (the dataset's single minutes value) a better or worse predictor of rating than ingredient count?

Findings

1. Ingredient/step count vs. rating: No meaningful relationship (r ≈ -0.001) — simple and complex recipes rate about the same.
2. Complexity-rating relationship by category: Correlations range from -0.03 to 0.06 across categories — direction and strength vary slightly, but none show a strong pattern.
3. Hidden gems: 15,304 recipes identified as highly-rated but under-reviewed (z-score > 0.75 within category, review count below the median).
4. Time vs. ingredients as predictors: Neither predicts rating well — minutes (r ≈ -0.023) is a slightly stronger (though still weak) predictor than ingredient count (r ≈ -0.001).

---

📁 Repository Structure

recipe-rescue/
├── data/
│   ├── raw/                # RAW_recipes.csv and RAW_interactions.csv (untouched, git-ignored)
│   └── processed/          # Cleaned, feature-engineered, analysis-ready data
│
├── notebooks/               # Exploratory analysis, step-by-step, one notebook per phase
│   ├── 01_scrape_explore.ipynb   # Loads and inspects the Kaggle CSVs, aggregates ratings
│   ├── 02_clean_features.ipynb
│   ├── 03_filtering_questions.ipynb
│   ├── 04_pivot_complexity_rating.ipynb
│   └── 05_numpy_stats_plots.ipynb
│
├── src/                      # Reusable code, imported by notebooks/scripts (not copy-pasted)
│   ├── __init__.py
|   |
│   ├── clean.py              # Parsing/feature functions (ingredient count, minutes cleaning,
│   │                         #   nutrition split, category assignment)
│   └── analysis.py           # Pivoting, correlation, z-score, bucketing helpers
│
│
├── outputs/            # Saved charts/plots (PNG)
│
├── .gitignore
├── requirements.txt
├── learning_skill.txt        # Skill-by-skill practice guide for this project
└── README.md

---

🛠️ Setup

git clone <your-repo-url>
cd recipe-rescue
python -m venv venv
source venv/bin/activate       # Windows: venv\Scripts\activate
pip install -r requirements.txt

Then download the dataset from Kaggle ("Food.com Recipes and Interactions") and place RAW_recipes.csv and RAW_interactions.csv in data/raw/.

The two files are large (roughly 295 MB and 350 MB), which is over GitHub's 100 MB file limit, so add data/raw/ to .gitignore and don't commit them.

---

🚀 Usage

1. Add the data — put RAW_recipes.csv and RAW_interactions.csv in data/raw/
2. Clean & engineer features (fills data/processed/)

   python -m src.clean

3. Explore in notebooks, in order, starting with notebooks/01_scrape_explore.ipynb

---

📊 Skills Practiced

Skill	Where
Python fundamentals (functions, control flow)	src/clean.py
Pandas Series/DataFrame basics, merging and aggregating (groupby, merge)	notebooks/01, notebooks/02, src/clean.py
Parsing messy columns (stringified lists, nutrition, tags)	notebooks/02, src/clean.py
Filtering & boolean indexing	notebooks/03
Reshaping / pivoting (pivot_table, melt)	notebooks/04, src/analysis.py
NumPy vectorized ops (correlation, z-scores, bucketing)	notebooks/05, src/analysis.py

---

📈 Sample Finding

Recipe complexity — ingredient count, step count, or time — has almost no bearing on how well a recipe is rated. What matters more is visibility: thousands of highly-rated recipes are sitting with very few reviews, effectively undiscovered.

---

📄 Data Sources

• Recipe data: the Kaggle Food.com Recipes and Interactions dataset
  • RAW_recipes.csv — one row per recipe: name, id, minutes, tags, nutrition, steps, ingredients, and more
  • RAW_interactions.csv — one row per rating/review (a rating of 0 means a review with no score, so it is excluded from averages)

---

🔭 Possible Extensions

• Add basic NLP on the review text (most common complaint words for low-rated recipes)
• Rank the top 10 hidden gems by rating and surface them in a dedicated view
• Build a small Streamlit app to filter recipes by "quick + highly rated"