Severing Spurious Correlations with Data Pruning

This repository contains the code for experiments in the ICLR 2025 Spotlight paper Severing Spurious Correlations with Data Pruning by Varun Mulchandani and Jung-Eun Kim.

More experiments to come soon!

CelebA Experiments

Please download and unzip the CelebA dataset from Liu et. al. 2015 (Aligned and Cropped Images) and move it to the CelebA/data/ sub-directory.

Sample Commands

# Create train, test and val splits.

bash create_init_dataset.sh

# Train model on unpruned dataset.

python3 generate_downstream.py --exp_name CelebA_sample_exp --dataset CelebA --n_epochs 25 --lr 1e-3 --weight_decay 1e-4 --method ERM --prune False
bash results/CelebA/CelebA_sample_exp/ERM_upweight_0_epochs_25_lr_0.001_weight_decay_0.0001/job.sh

# Save the original metadata with a different name.

bash change_meta.sh

# Prune the dataset. prune_oracle and prune_general take amount to be pruned between 0.1 - 0.97. Type of pruning can be "hardest" or "easiest".

python3 prune_oracle.py 0.5 hardest # Assuming access to spurious information. Reproduces results in Fig. 4.
python3 prune_general.py 0.5 hardest # Assuming access to no information. Reproduces results in Fig. 6. 
bash prune_metadata.sh

# Train on pruned dataset. Make sure to move the original training results elsewhere, or they'll get overwritten.

python3 generate_downstream.py --exp_name CelebA_sample_exp --dataset CelebA --n_epochs 25 --lr 1e-3 --weight_decay 1e-4 --method ERM --prune True
bash results/CelebA/CelebA_sample_exp/ERM_upweight_0_epochs_25_lr_0.001_weight_decay_0.0001/job.sh

CIFAR-10S Experiments

Sample Commands

# Compute Core Feature Difficulty

python3 ResNet20_main.py

# Store indices for samples with hardest and easiest core features.

python3 split.py

# Compute Spurious Misclassifications with easy spurious feature (S3 in text).

# Introduce synthetic spurious feature into samples with hard core features.
python3 ResNet20_main_spurious_easiest.py hardest 

# Introduce synthetic spurious feature into samples with easy core features.
python3 ResNet20_main_spurious_easiest.py easiest 

# Compute Spurious Misclassifications with hard spurious feature (S1 in text).

# Introduce synthetic spurious feature into samples with hard core features.
python3 ResNet20_main_spurious_hardest.py hardest 

# Introduce synthetic spurious feature into samples with easy core features.
python3 ResNet20_main_spurious_hardest.py easiest 

Please consider citing our paper if you find this repository useful in your work.

@inproceedings{
mulchandani2025ICLR,
title={Severing Spurious Correlations with Data Pruning},
author={Varun Mulchandani and Jung-Eun Kim},
booktitle={The Thirteenth International Conference on Learning Representations (ICLR)},
year={2025},
url={https://openreview.net/forum?id=Bk13Qfu8Ru}
}

References

Many of our experiments are built on top of implementations provided by the following papers:

1. Just Train Twice: Improving Group Robustness without Training Group Information