This repository provides a reference implementation of the algorithm described in the paper :
Importance Reweighting for Biquality Learning
Importance Reweighting for Biquality Learning (IRBL) is a meta algorithm that learns to reweight untrusted examples from a small trusted dataset to detect corrupted data and learn efficient classifiers under complex supervision deficiencies.
In order to run the experiments and generate results files and figures, run the following lines :
git clone https://github.com/pierrenodet/irbl.git
cd irbl
unzip data.zip
python src/main.py
Otherwise detailed results are provided for both supervisions deficiencies (NCAR and NNAR) in the results directory.
If you use IRBL in your research, please consider citing us :
@inproceedings{nodet2021importance,
title={Importance reweighting for biquality learning},
author={Nodet, Pierre and Lemaire, Vincent and Bondu, Alexis and Cornu{\'e}jols, Antoine and Ouorou, Adam},
booktitle={2021 International Joint Conference on Neural Networks (IJCNN)},
pages={1--8},
year={2021},
organization={IEEE}
}