Python - Recipe-Rating-Prediction-using-Machine-Learning-

Full machine learning pipeline including data cleaning, EDA, NLP (TF-IDF), feature engineering, model development, and evaluation for predicting recipe star ratings. Project Overview
This project analyzes a large dataset of recipe reviews with the goal of predicting star ratings (1–5) using machine learning techniques.
The dataset includes:

• likes_score
• dislike_index
• vote_ratio
• ranking_value
• review text
• user & recipe identifiers
• timestamps
• numeric recipe attributes (steps, ingredients, time)

The workflow covers the entire data science lifecycle:

• ✔ Data Cleaning
• ✔ Exploratory Data Analysis (EDA)
• ✔ Feature Engineering
• ✔ Text Vectorization (TF-IDF)
• ✔ Modeling (Logistic Regression & Random Forest)
• ✔ Evaluation & Feature Importance
• ✔ Insights & Recommendations

1. Data Cleaning
   Steps taken to prepare the dataset:

✔ Removing invalid and duplicate entries

• Ensured only 1–5 star ratings
• Removed duplicates to avoid bias

✔ Handling missing values

• Numeric columns → filled with median
• Text → filled with empty string
• Categorical IDs → filled with "unknown"

✔ Converting timestamps

• Extracted month, weekday, and hour

✔ Normalizing categories

• Lowercased and stripped whitespace for consistency

2. Exploratory Data Analysis (EDA) Key findings from descriptive analysis:

Rating Distribution

• Highly imbalanced dataset — majority of reviews are 5-star.

  Relationship between numeric features & ratings

• High-ingredient or long recipes show more rating variance.

  Review Length vs Rating

• Longer reviews generally correlate with higher ratings.

  Correlation Heatmap

• Strong correlation between ranking score and star rating.

• Weak correlation between most numeric features and stars → text matters most.

3. Feature Engineering

✔ Review Length
Calculated word count as a new feature.

✔ TF-IDF Vectorization
Converted review text into numerical vectors.
Helps identify strong positive & negative words:

• Positive: delicious, loved, good
• Negative: not, boring, bland

✔ Encoding Categorical Variables
Applied one-hot encoding to user_id and recipe_id.

✔ Scaling Numeric Features
Standardized continuous variables to improve model performance.

4. Model Development

Two models were built and compared:

Logistic Regression (Baseline Model)

• Tuned using GridSearchCV
• Best parameters:
  • C = 0.5
  • penalty = l2
• Benefits: Balanced performance
• Accuracy: 78%
• Stronger performance on minority classes than Random Forest

Random Forest (Advanced Model)

• Tuned using GridSearchCV
• Best parameters:
  • n_estimators = 200
  • max_depth = 20
• Learning non-linear relationships
• Accuracy: 82%
• Excellent on 5-star predictions, poor on minority classes

5. Evaluation

Logistic Regression

• Accuracy: 78%
• Macro F1: 0.45
• ROC AUC: 0.87
• More balanced across star categories

Random Forest

• Accuracy: 82%
• Macro F1: 0.34
• ROC AUC: 0.79
• Strong on 5-star class but weak on 1–3 stars

6. Feature Importance

Key Predictors:

• TF-IDF text features dominated importance
• Strong indicators:
  • Positive words: delicious, great, loved
  • Negative words: not, but, bland
• Review length was also an important contributor
• Numeric attributes had minimal impact

7. Insights & Interpretation

Data Distribution

• Highly imbalanced → difficult for models

• Oversampling or synthetic balancing is needed

  Model Performance

• Random Forest achieved higher accuracy

• Logistic Regression performed better across all classes

  Feature Importance

• Text features drive most of the predictive power

• Numeric features contribute less

8. Practical Implications
   The findings can be applied to:

• Recipe recommendation systems
• Improving user feedback understanding
• Identifying recipe strengths & weaknesses
• Content moderation & quality improvement
• Highlighting reliable user reviews

9. Recommendations for Improvement

10. Address Class Imbalance

• Use SMOTE, undersampling, or hybrid methods

2. Use Advanced NLP Models

• Replace TF-IDF with BERT or transformer embeddings

3. Collect More Metadata

• Include cuisine type, difficulty, demographics

4. Hybrid Modeling

• Combine ensemble methods with deep learning