🚖 NYC Taxi Fare Prediction - MLOps Project

Proyek MLOps end-to-end untuk memprediksi tarif taksi NYC menggunakan data NYC TLC Trip Record.

🔗 Live Demo: https://mlops-nyc-taxi-7eaf5edf4a58.herokuapp.com

---

📋 Fitur Utama

Komponen	Deskripsi
ML Model	Random Forest & Gradient Boosting dengan Optuna tuning
API	FastAPI dengan auto-documentation (Swagger)
Dashboard	HTML/CSS/JS dengan prediksi real-time
Monitoring	Drift detection untuk Distance & Target (Fare)
Registry	Blue/Green deployment dengan MLflow
Deployment	Docker + Heroku dengan auto-deploy dari GitHub

---

🏗️ Arsitektur

┌─────────────────────────────────────────────────────────┐
│                    BROWSER                               │
│   HTML Dashboard (Prediction + Monitoring)               │
└─────────────────────┬───────────────────────────────────┘
                      │ HTTP REST API
┌─────────────────────▼───────────────────────────────────┐
│                  FastAPI Server                          │
│  ├── POST /predict        → Prediksi tarif              │
│  ├── GET  /health         → Status server               │
│  ├── GET  /model/info     → Info model aktif            │
│  ├── GET  /monitoring/drift → Drift metrics             │
│  └── GET  /docs           → Swagger UI                  │
└─────────────────────┬───────────────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────────────┐
│              ML Model Layer                              │
│  ├── production_model.joblib (Model aktif)              │
│  ├── MLflow Registry (Version control)                  │
│  └── Reference Stats (Baseline untuk drift)             │
└─────────────────────────────────────────────────────────┘

---

📁 Struktur Project

mlops/
├── src/
│   ├── serving/
│   │   ├── api.py              # FastAPI server
│   │   ├── static/index.html   # Dashboard UI
│   │   └── reference_stats.json
│   ├── models/
│   │   ├── train.py            # Training pipeline
│   │   └── registry.py         # MLflow registry
│   └── features/
│       └── engineering.py      # Feature engineering
├── cli/
│   └── main.py                 # CLI commands
├── models/
│   └── production_model.joblib # Model production
├── Dockerfile                  # Docker config
├── heroku.yml                  # Heroku config
└── requirements.txt

---

🚀 Quick Start

Lokal (Development)

# Clone repo
git clone https://github.com/kemasverii/mlops-nyc-taxi.git
cd mlops-nyc-taxi

# Install dependencies
pip install -r requirements.txt

# Jalankan server
python cli/main.py serve start

# Buka browser: http://localhost:8000

Docker

# Build image
docker build -t mlops-api .

# Run container
docker run -p 8000:8000 mlops-api

---

🔧 CLI Commands

# Server
python cli/main.py serve start              # Start API server
python cli/main.py serve start -p 8080      # Custom port
python cli/main.py serve mlflow             # Start MLflow UI

# Model Registry
python cli/main.py registry status          # Lihat Blue/Green status
python cli/main.py registry list            # List semua versi
python cli/main.py registry promote 2       # Promote versi ke production
python cli/main.py registry rollback 1      # Rollback ke versi sebelumnya
python cli/main.py registry runs            # List MLflow runs
python cli/main.py registry register <run_id>  # Register model dari run

# Training
python cli/main.py train quick              # Quick training (no registry)
python cli/main.py train pipeline           # Production training + registry
python cli/main.py train compare-algos      # Compare semua algoritma
python cli/main.py train tune               # Hyperparameter tuning (Optuna)

# Data
python cli/main.py data --help              # Data operations

# Monitoring
python cli/main.py monitor --help           # Monitoring operations

# Model Testing
python cli/main.py model test <version>     # Test model tertentu
python cli/main.py model compare 1 2        # Compare dua versi

---

📊 Model Registry (Blue/Green)

Stage	Deskripsi
🔵 BLUE	Model production yang aktif
🟢 GREEN	Model staging untuk testing

Promote model dari GREEN ke BLUE:

python cli/main.py registry promote <version>

---

📈 Monitoring Dashboard

Dashboard menampilkan:

• Distance Drift - Pergeseran distribusi jarak trip
• Target Drift - Pergeseran prediksi fare
• Model Info - Nama & versi model aktif
• Charts - Visualisasi distribusi data

Drift Detection dengan Evidently AI

Sistem menggunakan Evidently AI untuk deteksi drift:

Komponen	Deskripsi
Reference Data	10,000 sample dari training data
Current Data	100 prediksi terakhir
Algoritma	Wasserstein Distance (numerik)
Threshold	0.3 (per fitur), 50% (dataset)

---

🔧 Feature Engineering

Input dari User (Web Form)

Fitur	Range	Deskripsi
PULocationID	Dropdown (263)	Zona pickup NYC
DOLocationID	Dropdown (263)	Zona dropoff NYC
passenger_count	1 - 6	Jumlah penumpang
pickup_hour	0 - 23	Jam pickup
pickup_dayofweek	0 - 6	Hari (0=Senin)
VendorID	Dropdown	Vendor taxi

Fitur yang Dihitung Otomatis

Fitur	Formula	Sumber Data
trip_distance	Lookup table	Rata-rata per rute (39,307)
is_weekend	1 if dayofweek >= 5 else 0	Dari input user
trip_duration_minutes	(distance / 11) * 60	11 mph = avg speed NYC
pickup_month	random(1-5)	Training data hanya Jan-Mei
hour_sin, hour_cos	Cyclical encoding	Pattern waktu circular
dow_sin, dow_cos	Cyclical encoding	Pattern hari circular
avg_speed_mph	distance / (duration/60)	= 11 mph
is_rush_hour	1 if 16 <= hour <= 19	Jam sibuk sore
same_location	1 if PU == DO	Dari input user

Nilai Fixed

Fitur	Nilai	Alasan
has_tolls	0	Simplifikasi untuk demo

---

📐 Route Distance Lookup

Cara Kerja

Sistem menggunakan pre-computed lookup table untuk estimasi jarak:

1. User pilih Pickup Location (dropdown 263 zona)
2. User pilih Dropoff Location (dropdown 263 zona)
3. Sistem query lookup table → auto-fill trip_distance

Data Lookup

Statistik	Nilai
Total Routes	39,307
Sumber	11M trips
Algoritma	Mean per route
Default (jika N/A)	3.0 miles

Contoh

Pickup: Midtown Center (161)
Dropoff: Upper East Side South (237)
→ Lookup: "161_237" = 1.07 miles

---

📊 Design Decisions

Mengapa AVG_SPEED = 11 mph?

Dihitung dari 11 juta trips seluruh NYC:

• Mean actual: 11.11 mph
• Dibulatkan: 11 mph

Mengapa pickup_month Random 1-5?

Training data hanya berisi bulan Januari-Mei:

• Jan: 1.9 juta trips
• Feb: 2.0 juta trips
• Mar: 2.4 juta trips
• Apr: 2.3 juta trips
• Mei: 2.5 juta trips

---

🌐 API Endpoints

Method	Endpoint	Deskripsi
GET	/	Dashboard HTML
GET	/health	Health check
POST	/predict	Prediksi tarif
GET	/model/info	Info model
GET	/zones	List 263 zona NYC
GET	/route-distance	Estimasi jarak rute
GET	/monitoring/drift	Drift metrics
GET	/docs	Swagger UI

Contoh Request

curl -X POST https://mlops-nyc-taxi-7eaf5edf4a58.herokuapp.com/predict \
  -H "Content-Type: application/json" \
  -d '{
    "trip_distance": 5.0,
    "passenger_count": 2,
    "pickup_hour": 14,
    "pickup_dayofweek": 2,
    "PULocationID": 161,
    "DOLocationID": 237,
    "pickup_month": 1
  }'

---

🚢 Deployment

Project ini di-deploy ke Heroku dengan auto-deploy dari GitHub:

1. Push ke GitHub → Heroku auto-rebuild
2. Zero-downtime deployment
3. Docker-based containerization

---

📝 Tech Stack

• Backend: FastAPI, Uvicorn
• ML: Scikit-learn, Pandas, NumPy
• Tracking: MLflow
• Frontend: HTML, CSS, JavaScript, Chart.js
• Deployment: Docker, Heroku
• CI/CD: GitHub (auto-deploy)

---

👤 Author

Kemas Veriandra Ramadhan

Ahmad Sahidin Akbar

Eli Dwi Putra Berema

Nisrina Nur Afifah

⁠Khaalishah Zuhrah Alyaa V.

---

📄 License

MIT License