Fair Tabular Diffusion

NOTE: The code for our method is in src, and the Python script for running experiments using our method is fairtabddpm_opt.py.

Setup

The PyTorch version we used in this project is 2.3.0+cu121, and you can install the required packages by running the following command:

conda create -n ai python=3.10
source activate ai
pip install -r requirements.txt
pip install dgl -f https://data.dgl.ai/wheels/torch-2.3/cu121/repo.html
pip install torch-scatter torch-sparse -f https://data.pyg.org/whl/torch-2.3.0+cu121.html

To download and preprocess the datasets, run the following command:

python build.py

Running Experiments

Under the root directory, run the following commands to reproduce the results of our method:

# run experiments for our method
bash fairtabddpm.sh

To reproduce the results of baseline methods, run the following commands:

# go to baselines directory
cd baselines
# run experiments for baselines
bash codi.sh
bash fairsmote.sh
bash fairtabgan.sh
bash goggle.sh
bash great.sh
bash smote.sh
bash stasy.sh
bash tabddpm.sh
bash tabsyn.sh

Benchmarks

Datasets

• Adult
• COMPASS
• German Credit
• Bank Marketing

Baselines

The baseline methods we used in this project are as follows (sorted alphabetically):

• CoDi
• Goggle
• GReaT
• SMOTE
• STaSy
• TabDDPM
• TabSyn
• Fair Class Balancing (FCB)
• FairTGAN

To Do

Avoid repeatition to improve the code quality:

• Replace exp_config['home'] by importing EXPS_PATH from constant.py in all running scripts
• Replace data_config['path'] by importing DB_PATH from constant.py in all running scripts
• Delete home of experiments and path of datasets in all config.toml files
• Add a new argument --method to optimization scripts and merge all optimization scripts into one
• Find commonly used functions in all running scripts and move them to utils.py

Organize the code:

• Move fairtabddpm.sh, fairtabddpm_run.py, fairtabddpm_opt.py to baseline directory and rename baseline directory to methods, and edit readme.md accordingly
• Move src/evaluate/metrics.py out to the root directory because it is specific to the project

Automate the experiments and evaluations:

• Refactor and reorganize assess/present.ipynb with functional programming
• Rewrite all the code in assess directory with functional programming

Correct the errors:

• The implementation of TabSyn in baselines is incorrect