puncc 0.7

Predictive UNcertainty Calibration and Conformalization Library

Puncc (short for Predictive uncertainty calibration and conformalization) is an open-source Python library. It seamlessly integrates a collection of state-of-the-art conformal prediction algorithms and associated techniques for diverse machine learning tasks, including regression, classification and anomaly detection. Puncc can be used with any predictive model to provide rigorous uncertainty estimations. Under data exchangeability (or i.i.d), the generated prediction sets are guaranteed to cover the true outputs within a user-defined error $\alpha$.

Documentation is available online.

📚 Table of contents

• 🐾 Installation
• 📖 Documentation
• 👨‍🎓 Tutorials
• 🚀 QuickStart
• 📚 Citation
• 💻 Contributing
• 🙏 Acknowledgments
• 👨‍💻 Creators
• 📝 License

🐾 Installation

puncc requires a version of python higher than 3.8 and several libraries including Scikit-learn and Numpy. It is recommended to install puncc in a virtual environment to not mess with your system's dependencies.

You can directly install the library using pip:

pip install git+https://github.com/deel-ai/puncc

📖 Documentation

For comprehensive documentation, we encourage you to visit the official documentation page.

👨‍🎓 Tutorials

We highly recommand following the introduction tutorials to get familiar with the library and its API:

• Introduction tutorial

• API tutorial

You can also familiarize yourself with the architecture of puncc to build more efficiently your own conformal prediction methods:

• Architecture overview

🚀 Quickstart

Conformal prediction enables to transform point predictions into interval predictions with high probability of coverage. The figure below shows the result of applying the split conformal algorithm on a linear regressor.

Many conformal prediction algorithms can easily be applied using puncc. The code snippet below shows the example of split conformal prediction wrapping a linear model, done in few lines of code:

from sklearn import linear_model
from deel.puncc.api.prediction import BasePredictor

# Load training data and test data
# ...

# Instanciate a linear regression model
# linear_model = ...


# Create a predictor to wrap the linear regression model defined earlier.
# This enables interoperability with different ML libraries.
# The argument `is_trained` is set to False to tell that the the linear model
# needs to be trained before the calibration.
lin_reg_predictor =  BasePredictor(linear_model, is_trained=False)

# Instanciate the split cp wrapper around the linear predictor.
split_cp = SplitCP(lin_reg_predictor)

# Fit model (as is_trained` is False) on the fit dataset and
# compute the residuals on the calibration dataset.
# The fit (resp. calibration) subset is randomly sampled from the training
# data and constitutes 80% (resp. 20%) of it (fit_ratio = 80%).
split_cp.fit(X_train, y_train, fit_ratio=.8)

# The predict returns the output of the linear model y_pred and
# the calibrated interval [y_pred_lower, y_pred_upper].
y_pred, y_pred_lower, y_pred_upper = split_cp.predict(X_test, alpha=alpha)

The library provides several metrics (deel.puncc.metrics) and plotting capabilities (deel.puncc.plotting) to evaluate and visualize the results of a conformal procedure. For a target error rate of $\alpha = 0.1$, the marginal coverage reached in this example on the test set is higher than $90$% (see Introduction tutorial):

90% Prediction Interval with Split Conformal Prediction.

More flexibility with the API

Puncc provides two ways of defining and using conformal prediction wrappers:

• A direct approach to run state-of-the-art conformal prediction procedures. This is what we used in the previous conformal regression example.
• Low-level API: a more flexible approach based of full customization of the prediction model, the choice of nonconformity scores and the split between fit and calibration datasets.

A quick comparison of both approaches is provided in the API tutorial for a regression problem.

📚 Citation

If you use our library for your work, please cite our paper:

@inproceedings{mendil2023puncc,
  title={PUNCC: a Python Library for Predictive Uncertainty Calibration and Conformalization},
  author={Mendil, Mouhcine and Mossina, Luca and Vigouroux, David},
  booktitle={Conformal and Probabilistic Prediction with Applications},
  pages={582--601},
  year={2023},
  organization={PMLR}
}

Puncc has been used to support the work presented in our COPA 2022 paper on conformal prediction for time series.

@inproceedings{mendil2022robust,
  title={Robust Gas Demand Forecasting With Conformal Prediction},
  author={Mendil, Mouhcine and Mossina, Luca and Nabhan, Marc and Pasini, Kevin},
  booktitle={Conformal and Probabilistic Prediction with Applications},
  pages={169--187},
  year={2022},
  organization={PMLR}
}

💻 Contributing

Contributions are welcome! Feel free to report an issue or open a pull request. Take a look at our guidelines here.

🙏 Acknowledgments

This project received funding from the French ”Investing for the Future – PIA3” program within the Artificial and Natural Intelligence Toulouse Institute (ANITI). The authors gratefully acknowledge the support of the DEEL project.

👨‍💻 Creators

Mouhcine MENDIL initially developed this library as a research tool, with assistance from Lucas MOSSINA. We have recently welcomed Joseba DALMAU to the team to help enhance puncc and work on the development of new features.

🔑 License

The package is released under MIT license.