Paper 1. Distributionally Robust Survival Analysis: A Novel Fairness Loss Without Demographics

Shu Hu, George H. Chen

Paper 2. Fairness in Survival Analysis with Distributionally Robust Optimization

Shu Hu, George H. Chen

This repository is the official implementation of our two papers "Distributionally Robust Survival Analysis: A Novel Fairness Loss Without Demographics" (ML4H, 2022) and "Fairness in Survival Analysis with Distributionally Robust Optimization" (JMLR, 2024).

---

Some of the codes are extracted from FairSurv and SODEN.

If you would like to use the SEER dataset, you should request access from https://seer.cancer.gov/data/.

Requirements

To install requirements:

pip install -r requirements.txt

Please also consider to install necessary packages from the SODEN.

How to run Dro-COX code

1. To run Dro-COX on FLC

For FLC (Linear),

python -u run_dro_cox.py --dataset FLC --model Linear --eps 0.15 --seed 7 > FLC_Linear_joint_dro.log 2>&1 & 

For FLC (MLP),

python -u run_dro_cox.py --dataset FLC --model MLP --eps 0.3 --seed 7 > FLC_MLP_joint_dro.log 2>&1 &  

2. To run Dro-COX on SUPPORT

For SUPPORT (Linear),

python -u run_dro_cox.py --dataset SUPPORT --model Linear --eps 0.15 --seed 7 > SUPPORT_Linear_joint_dro.log 2>&1 & 

For SUPPORT (MLP),

python -u run_dro_cox.py --dataset SUPPORT --model MLP --eps 0.5 --seed 7 > SUPPORT_MLP_joint_dro.log 2>&1 & 

3. To run Dro-COX on SEER

For SEER (Linear),

python -u run_dro_cox.py --dataset SEER --model Linear --eps 0.2 --seed 7 > SEER_Linear_joint_dro.log 2>&1 &

For SEER (MLP),

python -u run_dro_cox.py --dataset SEER --model MLP --eps 0.15 --seed 7 > SEER_MLP_joint_dro.log 2>&1 & 

How to run Dro-COX (SPLIT) code

1. To run Dro-COX (SPLIT) on FLC

For FLC (Linear),

python -u run_dro_cox_split.py --dataset FLC --model Linear --eps 0.1 --seed 7 > FLC_Linear_joint_dro_split.log 2>&1 & 

For FLC (MLP),

python -u run_dro_cox_split.py --dataset FLC --model MLP --lr 0.0001 --eps 0.05 --seed 7 > FLC_MLP_joint_dro_split.log 2>&1 &  

2. To run Dro-COX (SPLIT) on SUPPORT

For SUPPORT (Linear),

python -u run_dro_cox_split.py --dataset SUPPORT --model Linear --eps 0.15 --seed 7 > SUPPORT_Linear_joint_dro_split.log 2>&1 & 

For SUPPORT (MLP),

python -u run_dro_cox_split.py --dataset SUPPORT --model MLP --lr 0.0001 --eps 0.2 --seed 7 > SUPPORT_MLP_joint_dro_split.log 2>&1 & 

3. To run Dro-COX (SPLIT) on SEER

For SEER (Linear),

python -u run_dro_cox_split.py --dataset SEER --model Linear --eps 0.15 --seed 7 > SEER_Linear_joint_dro_split.log 2>&1 &

For SEER (MLP),

python -u run_dro_cox_split.py --dataset SEER --model MLP --lr 0.0001 --eps 0.2 --seed 7 > SEER_MLP_joint_dro_split.log 2>&1 & 

How to plot Figure 1 ( Dro-COX )

Setting dataset = 'FLC' in plot.py

For FLC,

python plot.py

Setting dataset = 'SUPPORT' in plot.py

For SUPPORT,

python plot.py

Setting dataset = 'SEER' in plot.py

For SEER,

python plot.py

How to run Dro-DeepHit code

1. To run Dro-DeepHit on FLC

For FLC (MLP),

python -u run_dro_deephit.py --dataset FLC --model MLP --eps 0.3 --seed 7 > FLC_MLP_joint_dro_deephit.log 2>&1 &  

2. To run Dro-DeepHit on SUPPORT

For SUPPORT (MLP),

python -u run_dro_deephit.py --dataset SUPPORT --model MLP --eps 0.5 --seed 7 > SUPPORT_MLP_joint_dro_deephit.log 2>&1 & 

3. To run Dro-DeepHit on SEER

For SEER (MLP),

python -u run_dro_deephit.py --dataset SEER --model MLP --eps 0.15 --seed 7 > SEER_MLP_joint_dro_deephit.log 2>&1 & 

How to run Dro-DeepHit (SPLIT) code

1. To run Dro-DeepHit (SPLIT) on FLC

For FLC (MLP),

python -u run_dro_deephit_split.py --dataset FLC --model MLP --lr 0.0001 --eps 0.05 --seed 7 > FLC_MLP_joint_dro_deephit_split.log 2>&1 &  

2. To run Dro-DeepHit (SPLIT) on SUPPORT

For SUPPORT (MLP),

python -u run_dro_deephit_split.py --dataset SUPPORT --model MLP --lr 0.0001 --eps 0.2 --seed 7 > SUPPORT_MLP_joint_dro_deephit_split.log 2>&1 & 

3. To run Dro-DeepHit (SPLIT) on SEER

For SEER (MLP),

python -u run_dro_deephit_split.py --dataset SEER --model MLP --lr 0.0001 --eps 0.2 --seed 7 > SEER_MLP_joint_dro_deephit_split.log 2>&1 & 

How to run Dro-SODEN code

Please first read the settings of the original SODEN code and then use our method.

1. To run Dro-SODEN on FLC

For FLC (MLP),

python -u ./SODEN/main_DRO_COX_ODE.py --dataset flc > FLC_MLP_joint_dro_SODEN.log 2>&1 &  

2. To run Dro-SODEN on SUPPORT

For SUPPORT (MLP),

python -u ./SODEN/main_DRO_COX_ODE.py --dataset support > SUPPORT_MLP_joint_dro_SODEN.log 2>&1 & 

3. To run Dro-SODEN on SEER

For SEER (MLP),

python -u ./SODEN/main_DRO_COX_ODE.py --dataset seer > SEER_MLP_joint_dro_SODEN.log 2>&1 & 

How to run the Exact DRO Cox code

1. To run the Exact DRO Cox on FLC

For FLC (MLP),

python -u run_dro_cox_full.py --dataset FLC > FLC_MLP_joint_dro_cox_full.log 2>&1 &  

2. To run the Exact DRO Cox on SUPPORT

For SUPPORT (MLP),

python -u run_dro_cox_full.py --dataset support > SUPPORT_MLP_joint_dro_cox_full.log 2>&1 & 

3. To run the Exact DRO Cox on SEER

For SEER (MLP),

python -u run_dro_cox_full.py --dataset seer > SEER_MLP_joint_dro_cox_full.log 2>&1 & 

Citation

Please kindly consider citing our paper in your publications.

@inproceedings{hu2022distributionally,
  title={Distributionally robust survival analysis: A novel fairness loss without demographics},
  author={Hu, Shu and Chen, George H},
  booktitle={Machine Learning for Health},
  pages={62--87},
  year={2022},
  organization={PMLR}
}