Skip to content

Latest commit

 

History

History
executable file
·
56 lines (50 loc) · 2.43 KB

README.md

File metadata and controls

executable file
·
56 lines (50 loc) · 2.43 KB

Incentivizing Collaboration in Machine Learning via Synthetic Data Rewards [AAAI-2022].

The code for our AAAI'22 paper "Incentivizing Collaboration in Machine Learning via Synthetic Data Rewards" which received an Oral presentation (4.26% acceptance rate).

Requirements

  1. Linux machine (experiments were run on Ubuntu 18.04.5 LTS and Ubuntu 20.04.2 LTS)
  2. NVIDIA GPU with CUDA 11.0 (experiments were run with NVIDIA GeForce GTX 1080 Ti and NVIDIA TITAN V)
  3. Anaconda (alternatively, you may install the packages in environment.yml manually)

Setup

In the main CGM directory,

  1. Run the following script to download the datasets.
bash download.sh
  1. Run the following command to install the required Python packages into a new environment named CGM using Anaconda.
conda env create -f environment.yml

Running experiments

In the main CGM directory,

  1. Change current environment to the CGM environment.
conda activate CGM
  1. Run the desired experiment. Valid values for dataset are {creditratings, creditcard, mnist, cifar}, valid values for split are {equaldisjoint, unequal}, and valid values for inv_temp are any non-negative real number.
python cgm.py with ${dataset} split=${split} inv_temp=${inv_temp}
# Example: to run the experiment on the creditcard dataset with the equal disjoint split and inv_temp = 1
python cgm.py with creditcard split=equaldisjoint inv_temp=1
  1. To replicate the correlation metrics in the paper for any dataset and split, run the following Python script to compute and display the metrics after all experiments with inv_temp = {1, 2, 4, 8} have completed.
python metrics.py wih ${dataset}
  1. To replicate the downstream supervised learning experiments in the paper for any dataset and split, run the following Python script to compute and display the metrics after all experiments with inv_temp = {1, 2, 4, 8} have completed.
python supervised.py wih ${dataset}

License

This code is released under the MIT License.

Citing our paper

If you find our paper relevant or use our code in your research, please consider citing our paper:

@InProceedings{tay2021,
  title={Incentivizing collaboration in machine learning via synthetic data rewards},
  author={Sebastian Shenghong Tay, Xinyi Xu, Chuan Sheng Foo, Bryan Kian Hsiang Low},
  booktitle={Proc. AAAI},
  year={2022}
}