Cocktail Recommender System

Using Real & Synthetic Data for Recommendations Evaluation

Using Machine Learning to Suggest Cocktails

Dharti Seagraves

Date: May 22, 2024

---

Problem Statement

• Objective: Create a cocktail recommender system using cocktail features and user ratings data
• Importance:
  • Personalized recommendations can enhance user experience and satisfaction in cocktail selection
  • Helps users discover new cocktails tailored to their tastes
  • Can help generate revenue for bars and restaurants
• Motivation: Improve customer engagement and satisfaction in the beverage industry

---

Dataset

• Cocktail Data Source: Web scraped from CocktailViz
  • 48 cocktails, 22 features
  • Details on ingredients, alcohol content, and flavor profiles
• User Ratings Data Source: Collected via a Google Form survey
  • Included user preferences and ratings on various cocktails
• Dataset Size:
  • Real Data: 115 records (actual user ratings)
  • Synthetic Data: 10,000 records (generated to compare with a larger dataset)

---

Assumptions and Hypotheses

• Assumptions:
  • User preferences are consistent over time
  • Survey bias does not exist
  • Synthetic data is a reliable approximation of real user behavior
• Hypotheses:
  • Feature importance aligns with user preferences
  • Synthetic data will not replicate real user data because it lacks real-world patterns (used primarily for the quantity of records)

---

Exploratory Data Analysis

• Correlation Matrix:
  • No multi-collinearity
  • 117 features after Feature Scaling
• Distribution Analysis:
  • Even distribution of liquors and alcohol content
• 3D Distribution of Users, Cocktails, and Ratings:
  • Synthetically generated user ratings:
    • Train/test split: 7500/2500
    • Unique Users: 2k

---

Feature Engineering & Transformations

• Feature Engineering: None conducted
• Transformations:
  • Normalization, encoding, and scaling techniques applied to improve model performance
  • Standard Scaling numerical values:
    • ['Alcohol', 'Base Wine Amount', 'Salty', 'Savory', 'Sour', 'Bitter', 'Sweet', 'Spicy']
  • One-hot encoding categorical values:
    • ['Category', 'Making', 'Base Wine', 'Liquor', 'Liquor Amount', 'Juice', 'Juice Amount', 'Spice', 'Spice Amount', 'Soda', 'Soda Amount', 'Others', 'Taste', 'Type of Glass']
• Feature Importance:
  • Sweet: 0.000569
  • Juice_Pineapple: 0.000374
  • Alcohol: 0.000354
  • Liquor_Green Mint: 0.000147
  • Base Wine_Rum: 0.000140

---

Model Development

• Approaches:
  • Item-based Collaborative Filtering: Recommends items similar to those the user liked
  • User-based Collaborative Filtering: Recommends items liked by similar users
  • Content-based Filtering: Recommends items similar to those the user has liked based on features
  • SVD-based Collaborative Filtering: Uses matrix factorization to predict user preferences
• Overfitting and Underfitting Checks:
  • Learning Curves: Analyzed to determine if models are learning effectively

---

User Profiles (Item-Based)

• User 2's Top-Rated Cocktails:
  • Tom Collins (5)
  • Gin & Tonic (4)
  • Mojito (4)
  • Black Russian (4)
  • Manhattan (3)
  • Screwdriver (3)
  • Cosmopolitan (1)
  • Bloody Mary (1)
  • Tequila Martini (1)
• Recommendations:
  • Blue Hawaii: Similar to Tom Collins
  • Vodka & Tonic: Similar to Gin & Tonic
  • Martini
• Similarity Insights:
  • User 38 rated both Tom Collins and Blue Hawaii highly

---

User Profiles (User-Based)

• User 2's Top-Rated Cocktails:
  • Tom Collins (5)
  • Gin & Tonic (4)
  • Mojito (4)
  • Black Russian (4)
  • Manhattan (3)
  • Screwdriver (3)
  • Cosmopolitan (1)
  • Bloody Mary (1)
  • Tequila Martini (1)
• User 38's Top-Rated Cocktails:
  • Gin & Tonic (5)
  • Moscow Mule (5)
  • Bloody Mary (5)
  • Margarita (5)
  • Blue Hawaii (4)
  • Tom Collins (4)
  • Long Island Ice Tea (4)
  • Manhattan (4)
  • Tequila Sunset (3)
  • Vodka & Tonic (3)
  • Screwdriver (1)
• Recommendations:
  • Martini
  • Moscow Mule
  • Long Island Ice Tea
• Similarity Insights:
  • User 2 is most similar to User 38 based on the similarity matrix

---

User Profiles (Content-Based)

• User 2's Top-Rated Cocktails:
  • Tom Collins (5)
  • Gin & Tonic (4)
  • Mojito (4)
  • Black Russian (4)
  • Manhattan (3)
  • Screwdriver (3)
  • Cosmopolitan (1)
  • Bloody Mary (1)
  • Tequila Martini (1)
• Recommendations:
  • Horse’s Neck
  • Moscow Mule
  • Vodka & Tonic
• Similarity Insights:
  • Based on Category, Making, and Taste similarities

---

User Profiles (SVD)

• User 2's Top-Rated Cocktails:
  • Tom Collins (5)
  • Gin & Tonic (4)
  • Mojito (4)
  • Black Russian (4)
  • Manhattan (3)
  • Screwdriver (3)
  • Cosmopolitan (1)
  • Bloody Mary (1)
  • Tequila Martini (1)
• Recommendations:
  • Margarita
  • Old Fashioned
  • Moscow Mule
• SVD Insights:
  • Leverages matrix factorization to reveal latent relationships between users and items

---

Model Evaluation

• Visualization:
  • Learning curves for the SVD model on both real and synthetic datasets for 20 epochs
• Metrics:
  • Real data - Train RMSE: 1.31 Test RMSE: 1.17
  • Synthetic data - Train RMSE: 0.97 Test RMSE: 0.97
• Observations:
  • Real Data: The training RMSE steadily decreases, indicating learning. The testing RMSE remains stable with a slight increase towards the end, indicating potential overfitting.
  • Synthetic Data: The synthetic data starts with similar RMSE values. The test RMSE has a steeper curve, indicating quicker learning and good generalization.

---

Results & Learnings

• Summary of RMSE Values:
  • Item-Based Collaborative Filtering RMSE: 3.86
  • User-Based Collaborative Filtering RMSE: 3.83
  • Content-Based Filtering RMSE: 2.96
  • SVD-Based Collaborative Filtering RMSE: 1.13
• Conclusion:
  • The best model for the given real dataset was the SVD model with the lowest RMSE.
  • The SVD model provides the closest recommendations to the user’s actual ratings.
  • Latent factors in SVD influence user preferences for more accurate recommendations.
  • The SVD model is recommended for making cocktail recommendations.

---

Future Work

• Increase the Volume of Data:
  • Collect more real user ratings to improve model accuracy.
  • Aim for at least 1000 real user ratings to better understand user preferences.
• Reduce User Bias:
  • Randomly showcase a few cocktails for users to rate versus allowing users to choose.
• Explore Advanced Algorithms:
  • Regularization and Hyperparameter Tuning.
  • Neural Collaborative Filtering.
  • Hybrid Models.
• Incorporate Time-Series Elements:
  • Track how user preferences change over time.
• A/B Testing and Continuous Improvement:
  • Implement A/B testing to evaluate model performance and improve iteratively.
• Expand Cocktail Selection:
  • Include a more diverse range of cocktails, including non-alcoholic options.

---

THANK YOU!