🍽️ ML-Powered Restaurant Recommendation System

An end-to-end machine learning recommendation system designed to reduce decision fatigue on food delivery platforms by 40%

💻Live Demo | 📓 Documentation | 📝 Full PRD | 🐞 Report Bug

▶️ Try the Live Demo:

---

📸 Project Demo

Personalized Recommendations with Explanations

Interactive Dashboard

Cold Start Onboarding Flow

---

📋 Project Overview

Problem Statement

Food delivery users face significant friction in the ordering process due to overwhelming restaurant choices and lack of contextual prioritization. This increases:

• ⏱️ Decision Fatigue: Users spend 8-12 minutes browsing 200+ restaurant options and get overwhelmed
• 🛒 Cart Abandonment: ~30% of users add items but don't complete checkout
• 🔄 Low Discovery: 65% of orders are repeat orders from the same 3-5 restaurants
• 💸 Revenue Loss: High-quality restaurants with availability go undiscovered

Root Cause: Current "Sort by: Distance/Rating/Delivery Time" approach is too generic and doesn't consider:

• User's historical preferences (cuisine, price, dietary needs)
• Contextual factors (time of day, weather, occasion)
• Real-time constraints (restaurant availability, delivery capacity)

---

✨ Solution Overview

An ML-powered hybrid recommendation system specifically for the home feed that combines:

1. Collaborative Filtering (40% weight)

• Learns from similar users' preferences
• "Users like you ordered from these restaurants"
• Enables discovery of unexpected matches

2. Content-Based Filtering (35% weight)

• Matches restaurant attributes to user profile
• Considers cuisine, price, rating, dietary restrictions
• Works even for new users (cold start handling)

3. Contextual Factors (25% weight)

• Time of Day: Breakfast → South Indian/Cafe, Dinner → Biryani/Chinese
• Weather: Rainy → Comfort food, Hot → Beverages/Desserts
• Distance: Exponential decay penalty for far restaurants
• Popularity: Slight boost for trending restaurants

Hybrid Score Calculation:

Final Score = 0.40 × CF_Score + 0.35 × CB_Score + 0.25 × Context_Score

---

⭐ Success Metrics

Primary Metric (Hero Metric):

• Time to Order: Reduce by 40% (10 minutes → 6 minutes)

Secondary Metrics:

• Order conversion rate: +15% improvement
• Restaurant discovery: 2+ new restaurants per user per month
• Repeat order rate: Decrease from 65% to 55%

Guardrail Metrics:

• Average delivery time: ≤38 minutes
• Order cancellation rate: ≤6%
• User dissatisfaction: ≤10%

---

🚀 Key Features

🎯 Personalized Recommendations Top-10 tailored to each user Context-aware (time, weather, location) Excludes closed restaurants <2 second response time 🧊 Cold Start Handling 3-question onboarding (<30 seconds) Content-based fallback strategy Works from first order 💡 Explainable AI Clear reason for each recommendation Examples: "You've ordered South Indian 3 times" Builds trust before ordering

🎨 Diversity Constraints Max 3 restaurants per cuisine Prevents filter bubble Balances familiarity (40%) with discovery (60%) 📊 Comprehensive Evaluation Precision@K, Recall@K, Hit Rate, NDCG Discovery metrics (diversity, novelty, coverage) Temporal train-test split 🔧 Production-Ready Fallback strategies for failures Edge case handling Complete documentation

---

📁 Project Structure

ml-restaurant-recommendations/
│
├── data/                                   # All datasets
│   ├── synthetic/                          # Generated data
│   │   ├── users.csv                       # 50K users
│   │   ├── restaurants.csv                 # 500 restaurants
│   │   └── orders.csv                      # 200K orders
│   └── processed/                          # Engineered features
│       ├── user_features.csv
│       ├── restaurant_features.csv
│       └── interaction_matrix.csv
│
├── src/                                    # Source code
│   ├── config.py                           # Configuration
│   ├── data_generator.py                   # Synthetic data creation
│   ├── feature_engineering.py              # Feature engineering
│   ├── collaborative_filtering.py          # CF model
│   ├── content_based_filtering.py          # CBF model
│   ├── hybrid_recommender.py               # Hybrid system
│   ├── explainability.py                   # Explanation engine
│   ├── cold_start_handler.py               # New user handling
│   └── evaluation.py                       # Model evaluation
│
├── app/                                    # Streamlit application
│   └── streamlit_app.py                    # Interactive demo
│
├── models/                                 # Saved models
│   ├── collaborative_model.pkl
│   ├── content_based_model.pkl
│   └── hybrid_model.pkl
│
├── prd/                                    # Product documentation
│   ├── ab_test_plan.md                     # A/B testing strategy
│   ├── edge_cases.md                       # Handling edge cases
│   └── restaurant_recommendations_prd.md   # Full PRD
│
├── scripts/                                # Utility scripts
│   └── train_models.py                     # Master training script
│
├── tests/                                  # Unit tests
│   ├── test_recommender.py
│   ├── test_cold_start.py
│   └── test_explainability.py
│
├── docs/                                   # Technical documentation
│    ├── architecture.md                    # System design and data flow
│    ├── methodology.md                     # ML approach and evaluation
│    └── lab_logbook.md                     # Development log
│
├── .gitignore                              # Git ignore patterns
├── requirements.txt                        # Dependencies
├── LICENSE                                 # MIT License
└── README.md                               # This file

---

⚡ Quick Start

Prerequisites

• Python 3.13 or higher
• pip package manager
• 4GB RAM minimum
• 500MB disk space

Installation

Step 1: Clone the repository

git clone https://github.com/iamAyushSaxena/ML-Restaurant-Recommendations.git
cd ml-restaurant-recommendations

Step 2: Setup environment

# Create virtual environment
python -m venv venv

# Activate virtual environment
source venv/bin/activate           # On MacOS
                                      # OR
venv\Scripts\activate              # On Windows

Step 3: Install dependencies

pip install -r requirements.txt

Step 4: Generate Data & Train Models

# Run the master training script to generate data and train all models
python scripts/train_models.py

This will:

1. Generate synthetic dataset (50K users, 500 restaurants, 200K orders)
2. Engineer features for users and restaurants
3. Train collaborative filtering model
4. Train content-based filtering model
5. Create hybrid recommendation system

Expected runtime: ~3-5 minutes

Step 5: Running the Demo

# Launch the interactive Streamlit application
streamlit run app/streamlit_app.py

The app will open in your browser at http://localhost:8501

---

🎮 Demo

🌐 Live Demo

👉 Try the Interactive Demo on Streamlit Cloud

---

🔬 Technical Approach

Model Architecture

User Request
    ↓
Extract Context (time, location, weather)
    ↓
    ├─────────────┬─────────────┬─────────────┐
    │             │             │             │
    ▼             ▼             ▼             ▼
Collaborative  Content-Based  Contextual   Business
Filtering      Filtering      Scoring      Rules
(40%)          (35%)          (25%)        
    │             │             │             │
    └─────────────┴─────────────┴─────────────┘
                    ↓
            Hybrid Ranker
                    ↓
        Apply Filters & Diversity
                    ↓
        Generate Explanations
                    ↓
        Return Top-N Recommendations

Algorithms Used

Collaborative Filtering:

• User-based CF with cosine similarity
• Top-30 similar users for recommendation
• Handles sparsity with sparse matrix representation

Content-Based Filtering:

• Weighted feature matching (cuisine 40%, price 25%, rating 20%, delivery 15%)
• StandardScaler normalization
• Dietary restriction hard filters

Contextual Scoring:

• Time-based cuisine boosting (1.3-1.5× multiplier)
• Weather-based adjustments
• Distance decay (exponential: e^(-dist/3km))

Hybrid Combination:

if user_order_count >= 3:
    final_score = 0.40*cf + 0.35*cb + 0.25*context
else:  # Cold start
    final_score = 0.75*cb + 0.25*context

---

📊 Evaluation Results

Evaluated on 100 test users with temporal train-test split:

Accuracy Metrics

Metric	@5	@10	@20	Interpretation
Precision	0.0842	0.0756	0.0621	7.6% of top-10 were ordered
Recall	0.1234	0.2145	0.3521	21.5% of actual orders in top-10
Hit Rate	0.3156	0.4823	0.6421	48.2% users ordered from top-10
NDCG	0.2134	0.2567	-	Good ranking quality

Discovery Metrics

Metric	Score	Interpretation
Diversity	0.7234	7+ cuisines in top-10 recommendations
Novelty	0.6421	64% are new restaurants for user
Coverage	0.4523	45% of catalog recommended across all users

Baseline Comparison

Approach	Hit Rate@10	Diversity	Trade-off Decision
Random	0.05	0.85	❌ Too low accuracy
Popular Only	0.32	0.23	❌ No personalization
Pure CF	0.51	0.58	⚠️ Better accuracy but lower diversity
Hybrid (Ours)	0.48	0.72	✅ Best balance

Decision Rationale: Traded 3% hit rate for 24% more diversity to prevent recommendation fatigue.

---

💡 Product Thinking

Key Design Decisions

1. Time-to-Order over CTR

❌ Could have optimized for: Click-through rate (common ML metric)
✅ Chose instead: Time-to-order

Why: CTR is a vanity metric. It doesn't address the user's real pain: decision fatigue. Time-to-order directly measures whether we're solving the problem.

PM Perspective:

"I could have optimized for CTR—that's what most ML projects do. But CTR doesn't address the user's pain point. A user clicking through 50 restaurants still takes 10 minutes to order. Time-to-order measures what actually matters: are we reducing decision fatigue?"

2. Explainability over Pure Accuracy

❌ Could have used: Deep learning (2-3% better accuracy)
✅ Chose instead: Simple models with clear explanations

Why: Food recommendations require trust. Users won't order from a restaurant they don't understand why it was suggested. Explainability isn't optional—it's core to adoption.

PM Perspective:

"I deliberately chose simpler models over deep learning. Why? Because food needs trust. Users won't order from a 'black box' recommendation. Every suggestion has a clear reason: 'You've ordered South Indian 3 times, rated 4.5/5 by 856 customers.' That's worth more than 2% accuracy."

3. Diversity Constraints over Pure Precision

❌ Could have shown: Top-10 all from user's favorite cuisine
✅ Chose instead: Max 3 restaurants per cuisine

Why: Pure precision creates a filter bubble. Long-term user satisfaction requires variety. Prevent recommendation fatigue.

PM Perspective:

"I enforce a max of 3 restaurants per cuisine in the top-10. This costs some precision but prevents the filter bubble. If I only showed North Indian because that's what you ordered before, you'd get bored fast. Discovery matters for retention."

Success Metrics Framework

Primary Metric (North Star):

• Time to Order: 40% reduction (10 min → 6 min)
• Why Primary: Directly addresses user pain point

Secondary Metrics:

• Order conversion rate: +15%
• Restaurant discovery: 2+ new restaurants/month
• Why Secondary: Important but not the core problem

Guardrail Metrics:

• Average delivery time: ≤38 minutes
• Order cancellation rate: ≤6%
• User dissatisfaction: ≤10%
• Why Guardrails: Prevent quality degradation while optimizing primary metric

Stakeholder Balance:

• Users: Want relevance + variety (conflicting!)
• Restaurants: Want visibility (but fair distribution)
• Platform: Wants revenue (but not at cost of trust)

Trade-offs Made:

1. Explainability > Accuracy: Users need reasons before ordering food
2. Diversity > Precision: Prevent recommendation fatigue
3. Speed > Perfection: 6-minute decision time is "good enough"

---

📚 Documentation

Product Documentation

• Complete PRD - Full product requirements document
• A/B Test Plan - Statistical testing methodology
• Edge Cases - Error handling and fallback strategies

Technical Documentation

• Methodology - ML approach, feature engineering, evaluation
• Architecture - System design and data flow
• Lab Logbook - Step-by-step development process

---

🧪 Testing

Run unit tests:

# Install pytest
pip install pytest pytest-cov

# Run all tests
pytest tests/ -v

# Run with coverage
pytest tests/ --cov=src --cov-report=html

# Run specific test file
pytest tests/test_recommender.py -v

---

🔮 Future Enhancements

Out of Scope (V1) - Noted in PRD

• ❌ Dynamic pricing optimization
• ❌ Restaurant commission strategies
• ❌ Courier assignment logic
• ❌ Long-term personalization (cross-month)
• ❌ Multi-city rollout strategy

Potential V2 Features

• Real-time availability filtering
• Group ordering recommendations
• Dietary restriction hard filters (allergies)
• A/B test framework implementation
• Multi-armed bandit for exploration

---

🤝 Contributing

Contributions are welcome! This is a portfolio project, but I'm happy to accept improvements.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

---

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.

TL;DR: You can freely use, modify, and distribute this project, even commercially, as long as you include the original license.

---

📞 Contact & Connect

👤Author: Ayush Saxena

• 🔗 LinkedIn: Ayush Saxena
• 🐙 GitHub: iamAyushSaxena
• 📧 Email: aysaxena8880@gmail.com

---

🙏 Acknowledgments

• Problem Inspiration: Real-world challenges in food delivery personalization
• Educational Value: Demonstrates end-to-end ML product development
• Portfolio Purpose: Showcases product thinking + technical execution for PM roles

---

⭐ Star This Repo

If you found this project helpful or impressive, please consider:

• ⭐ Starring the repository (helps others discover it)
• 🔄 Sharing on LinkedIn (tag me!)
• 💬 Providing feedback (open an issue with suggestions)
• 🍴 Forking for your own research (with attribution)

---

⭐ Star this repository if you found it valuable!

💬 Questions? Open an issue

🤝 Feedback? Start a discussion

---

Built with product thinking😎, not just algorithms!