Metabu - Learning meta-features

This is the Official code for ICLR 2022 paper "Learning meta-features for AutoML", Herilalaina Rakotoarison and Louisot Milijaona and Andry Rasoanaivo and Michèle Sebag and Marc Schoenauer.

This repository is still under active developement.

Installation

Install with pip:

pip install -r requirements.txt
python setup.py install

Use Singularity:

• build local container with definition file env/metabu.def.
• fetch remote container from (coming soon).

Usage

Simple to use:

from metabu import Metabu

basic_representations = pd.read_csv(...)
target_representations = pd.read_csv(...)
metabu = Metabu()
metabu.train(basic_reprs=basic_representations,
             target_reprs=target_representations,
             column_id="task_id")
metabu.predict(basic_reprs=basic_representations)
metabu.get_importances()

Try: cd examples; python metabu_adaboost.py

Feel free to create an issue if you have questions.

Experiments

Script to reproduce experiments will be available under the experiments branch.

Credits

• We use the implementation of the ICML 2020 work "Learning Autoencoders with Relational Regularization" [https://arxiv.org/pdf/2002.02913.pdf] to compute the Fused-Gromov-Wasserstein distance.
• We also grateful to the maintainers and contributors of the Python libraries in requirements.txt.

Cite Metabu

@inproceedings{rakotoarison2022learning,
    title       = {Learning meta-features for Auto{ML}},
    author      = {Herilalaina Rakotoarison and Louisot Milijaona and Andry Rasoanaivo and Michele Sebag and Marc Schoenauer},
    booktitle   = {International Conference on Learning Representations},
    year        = {2022},
    url         = {https://openreview.net/forum?id=DTkEfj0Ygb8}
}