woningwaardering-ml

Woningwaardering ML

A gradient boosting regression model for predicting Dutch housing valuation points (woningwaardering points) based on property characteristics.

This project is done for learning and educational purposes. It is not intended to be used for real-world applications.

Model Details

Model Description

This is a gradient boosting regression model that predicts Dutch housing valuation points (woningwaardering points) based on property characteristics. The model uses XGBoost's XGBRegressor with hyperparameter tuning.

• Developed by: Tomer Gabay
• Model type: XGBoost Regressor
• Language: Python
• License: MIT

Intended Use

Primary Use Cases

• Researching the impact of housing characteristics on woningwaardering points
• Predicting housing valuation points for properties in the Netherlands
• Real estate valuation assistance
• Housing market analysis

Out-of-Scope Use Cases

• Legal or official property valuation (this is a predictive model, not an official assessment)
• Properties outside the Netherlands
• Properties with characteristics significantly different from the training data
• Non-self-contained dwellings

Quick Start

Installation

Using uv (recommended)

# Install uv if you haven't already
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install dependencies
uv pip install -e .

Using pip

# Install as package (recommended)
pip install -e .

# Or just install dependencies
pip install -r requirements.txt

Training the Model

1. Open and run the train.ipynb notebook
2. The notebook includes:
   • Data loading from Hugging Face
   • Data preprocessing and feature engineering
   • Model training with hyperparameter tuning (RandomizedSearchCV with 5-fold cross-validation)
   • Model evaluation and metrics
   • Feature importance analysis
   • Visualizations
   • Model saving

Using a Trained Model

Option 1: Using the High-Level API (Recommended)

from woningwaardering_ml import WoningwaarderingPredictor
import pandas as pd

# Initialize predictor (loads model and preprocessing components)
predictor = WoningwaarderingPredictor()

# Prepare your data
new_property = pd.DataFrame({
    'year_built': [1990],
    'number_of_rooms': [5],
    'number_of_bedrooms': [3],
    'indoor_area': [120],
    'outdoor_area': [50],
    'energy_label': ['B'],
    'property_value': [350000]
})

# Make prediction
predictions = predictor.predict(new_property)
print(f"Predicted woningwaardering points: {predictions.iloc[0]}")

# Batch prediction
multiple_properties = pd.DataFrame({
    'year_built': [1990, 2010, 1985],
    'number_of_rooms': [5, 4, 6],
    'number_of_bedrooms': [3, 2, 4],
    'indoor_area': [120, 90, 150],
    'outdoor_area': [50, 30, 80],
    'energy_label': ['B', 'A', 'C'],
    'property_value': [350000, 280000, 420000]
})

batch_predictions = predictor.predict(multiple_properties)
print(batch_predictions)

Option 2: Using Lower-Level Functions

from woningwaardering_ml import load_model, load_preprocessing_components, preprocess_data
import pandas as pd

# Load model and preprocessing components
model = load_model('woningwaardering_model.json')
energy_label_encoder, feature_cols = load_preprocessing_components()

# Prepare your data
new_property = pd.DataFrame({
    'year_built': [1990],
    'number_of_rooms': [5],
    'number_of_bedrooms': [3],
    'indoor_area': [120],
    'outdoor_area': [50],
    'energy_label': ['B'],
    'property_value': [350000]
})

# Preprocess data
new_property_processed, _ = preprocess_data(
    new_property, 
    energy_label_encoder=energy_label_encoder
)

# Select features in the same order as training
X_new = new_property_processed[feature_cols]

# Make prediction
prediction = model.predict(X_new)
print(f"Predicted woningwaardering points: {prediction[0]}")

Option 3: Using Command Line Interface

# Interactive mode
python main.py

# With JSON input
python main.py --property '{"year_built":1990,"number_of_rooms":5,"number_of_bedrooms":3,"indoor_area":120,"outdoor_area":50,"energy_label":"B","property_value":350000}'

# With input file
python main.py --input properties.json --output predictions.csv

Required Input Features

The model requires the following features:

• year_built (int): Year the property was built
• number_of_bedrooms (int): Number of bedrooms.
• indoor_area (float): Total indoor area in m², including storage space and garage.
• outdoor_area (float): Total outdoor area in m².
• energy_label (str): Most recent valid energy label. One of: G, F, E, D, C, B, A, A+, A++, A+++, or A++++ (ordered worst to best). A missing energy label is treated as the worst label (G).
• property_value (float): Most recent official property value (WOZ waarde), as issued by the government, in euros.

Project Structure

woningwaardering-ml/
├── woningwaardering_ml/        # Python package
│   ├── __init__.py             # Package initialization
│   ├── preprocessing.py        # Data preprocessing utilities
│   ├── model.py                # Model loading utilities
│   └── inference.py            # High-level inference API
├── train.ipynb                 # Main training and evaluation notebook
├── main.py                     # CLI entry point for inference
├── push_to_hub.py              # Script to push model to Hugging Face Hub
├── pyproject.toml              # Project dependencies (uv)
├── requirements.txt            # Alternative requirements file (pip)
├── PUBLISHING_GUIDE.md         # Guide for publishing to Hugging Face
├── README.md                   # This file
├── LICENSE                     # MIT License
├── .gitignore                  # Git ignore rules
│
├── Generated files (after training):
├── woningwaardering_model.json  # Trained model
├── energy_label_encoder.pkl    # Energy label encoder
├── feature_columns.json        # Feature column names (order matters)
└── model_metadata.json         # Model metadata and hyperparameters

Training Data

Dataset

• Name: woonstadrotterdam/woningwaarderingen-collab
• Source: Hugging Face Datasets
• Splits: Train (80%), Validation (10%), Test (10%)
• Organizations: Woonstad Rotterdam, Thuisvester and Beveland Wonen
• Total samples: 48,550
• Training samples: 43,694 (train + val, used cross-validation)

For detailed dataset information, see the dataset card.

Data Preprocessing

• Removed description, single_family_home, number_of_rooms, organization columns
• Ordinal encoding for energy_label (G to A++++) → energy_label_encoded
• Final model uses 6 features:
  • year_built
  • number_of_bedrooms
  • indoor_area
  • outdoor_area
  • property_value
  • energy_label_encoded

Performance

Metrics

• Test MAE: 2.723 points
• Test MAPE: 1.662%

Evaluation Results

The model was evaluated on a held-out test set of 4,856 samples.

Model Architecture

Hyperparameters

• Algorithm: XGBRegressor
• Hyperparameter Tuning: RandomizedSearchCV with 5-fold cross-validation
• Reproducibility: Random seed = 42 (set for NumPy, Python random, and all models including XGBoost)
• Features: 6 features (year_built, number_of_bedrooms, indoor_area, outdoor_area, property_value, energy_label_encoded)
• Best parameters found:
  • n_estimators: 2000
  • max_depth: 9
  • learning_rate: 0.03
  • subsample: 0.8
  • colsample_bytree: 0.9
  • min_child_weight: 1
  • reg_alpha: 0
  • reg_lambda: 1

Changelog

Version 2.202607.0 (Oct 2026)

• Dataset update: Retrained on the July 2026 update of woonstadrotterdam/woningwaarderingen-collab, which grows from 27,496 to 48,550 dwellings.
  • Adds Beveland Wonen as a third organization.
  • The Woonstad Rotterdam data is rebuilt from scratch (16,927 → 30,588 dwellings). Its points now include shared spaces, kitchen and sanitary amenities, and shared spaces are excluded from indoor_area and outdoor_area.
  • The dataset no longer contains rows with a missing energy label, so a missing energy_label is now predicted as the worst label (G).
• Performance: Test MAE is 2.723 points and test MAPE is 1.662%, measured on the new test set of 4,856 dwellings. These are not comparable to the 2.274 points and 1.348% of version 1.1.0, which were measured on the previous test set: the new Woonstad Rotterdam points depend more on amenities that are not among the model features. The previous model scores an MAE of 6.06 points on the new test set.
• Hyperparameters: The search was rerun and selected the same parameters as version 1.1.0.
• Versioning: The version number now follows the dataset release it was trained on.

Version 1.1.0 (Jan 2026)

• Model upgrade: Migrated from scikit-learn's GradientBoostingRegressor to XGBoost's XGBRegressor
• Performance improvements:
  • Test MAE reduced from 2.97 to 2.274 points
  • Test MAPE reduced from 1.91% to 1.348%
• Training optimization: Switched from GridSearchCV to RandomizedSearchCV for faster hyperparameter tuning
• Dataset update:
  • Updated to use woonstadrotterdam/woningwaarderingen-collab dataset, which is a larger and more diverse dataset than the original dataset.
  • Combined train and validation sets since we use cross-validation during hyperparameter tuning.
• Dependencies: Added xgboost>=2.0.0 requirement

Version 1.0.0 (Oct 2025)

• Initial release with GradientBoostingRegressor
• Test MAE: 2.97 points
• Test MAPE: 1.91%
• Based on https://huggingface.co/datasets/woonstadrotterdam/woningwaarderingen

Development

Requirements

• Python >= 3.13
• See pyproject.toml or requirements.txt for dependencies

Reproducibility

The model uses a fixed random seed (42) for reproducibility. This is set for:

• NumPy random number generation
• Python's random module
• All models including XGBoost

Running Tests

Currently, the main training and evaluation is done in the train.ipynb notebook. The notebook includes:

• Train/validation/test split evaluation
• Cross-validation during hyperparameter tuning
• Feature importance analysis

Or use the notebook cell to push directly.

Citation

If you use this model, please cite:

@software{woningwaardering_ml,
  title = {Woningwaardering ML Model},
  author = {Gabay, Tomer},
  organization = {Woonstad Rotterdam},
  year = {2025},
  url = {https://huggingface.co/woonstadrotterdam/woningwaardering-ml}
}

License

MIT

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Contact

• Email: [email protected]
• Organization: Woonstad Rotterdam

Limitations

1. Geographic Scope: Trained on Rotterdam, Brabant and Zeeland data, may not generalize to other regions
2. Temporal Limitations:
   • Based on historical data, may not reflect future regulatory changes
3. Dwelling Type: Only trained on self-contained dwellings; will not work for non-self-contained dwellings
4. Data Quality: Performance depends on quality and completeness of input features
5. Dataset Bias: The dataset is biased towards social housing dwellings and may have less representation of higher-end dwellings

For more information about the dataset, see the dataset card.