Hotel Reservation Prediction

An end-to-end MLOps pipeline for predicting hotel reservation cancellations. This project leverages LightGBM for classification, Optuna for hyperparameter optimization, and MLflow for experiment tracking and model management.

🚀 Overview

The goal of this project is to predict the booking_status (Canceled vs. Not Canceled) of hotel reservations. This enables:

• Revenue Management: Better forecasting of cancellations.
• Targeted Marketing: Identifying customers likely to cancel and offering incentives.
• Operational Efficiency: Optimizing room allocation.

🛠️ Tech Stack

• Language: Python 3.12+
• Dependency Management: uv
• Model: LightGBM
• Hyperparameter Tuning: Optuna
• Experiment Tracking: MLflow
• Data Storage: Google Cloud Storage (GCP)
• Data Format: Parquet (for intermediate storage)
• Preprocessing: Scikit-learn, Imbalanced-learn (SMOTE)

📁 Project Structure

├── artifacts/              # Data and model artifacts
├── config/                 # Configuration files (YAML and Python)
├── logs/                   # Application logs
├── mlruns/                 # MLflow tracking data
├── pipeline/               # Pipeline script
    └── pipeline.py
├── src/                    # Source code for the pipeline
│   ├── data_ingestion.py   # GCP download and data splitting
│   ├── data_preprocessing.py # Feature engineering and balancing
│   ├── model_training.py    # Training with Optuna and MLflow
│   ├── logger.py           # Logging configuration
│   └── custom_exceptions.py # Custom error handling
├── utils/                  # Shared utility functions
├── pyproject.toml          # Project dependencies and metadata
└── README.md               # Project documentation

⚙️ Setup

1. Prerequisites

• uv installed.
• GCP Service Account Key (JSON) with access to the configured bucket.

2. Environment Variables

Create a .env file in the root directory:

GOOGLE_APPLICATION_CREDENTIALS="path/to/your/service-account-key.json"

3. Install Dependencies

uv sync

🛤️ Pipeline Execution

The pipeline is split into three main stages. You can run them sequentially or with the pipeline script:

Sequential Execution

1. Data Ingestion

Downloads the raw dataset from GCP and performs an initial train-test split.

uv run python -m src.data_ingestion

2. Data Preprocessing

Handles categorical encoding, log transformation for skewed features, data balancing using SMOTE, and feature selection using a Random Forest importance ranking.

uv run python -m src.data_preprocessing

3. Model Training

Trains a LightGBM model with:

• Hyperparameter Tuning: 100 trials using Optuna.
• MLflow Tracking: Automatically logs parameters, metrics (Accuracy, Precision, Recall, F1), and artifacts (datasets and the final model).

uv run python -m src.model_training

Pipeline Execution

uv run pipeline.pipeline

📊 Experiment Tracking with MLflow

To visualize the training results, metrics, and compare different runs:

1. Start the MLflow UI:

   mlflow ui

2. Open http://localhost:5000 in your browser.

📈 Dataset

The project uses the Hotel Reservations Classification Dataset from Kaggle.

• Target: booking_status
• Features: Includes lead time, average price per room, special requests, and more.