Fair yet Asymptotically Equal Collaborative Learning [ICML 2023]

This repository is the official implementation of the following paper accepted by the Fortieth International Conference on Machine Learning (ICML) 2023:

@inproceedings{Lin2023,
  author = {Xiaoqiang Lin, Xinyi Xu, See-Kiong Ng, Chuan-Sheng Foo, Bryan Kian Hsiang Low},
  title = {Fair yet Asymptotically Equal Collaborative Learning},
  booktitle = {Proceedings of the 40th International Conference on Machine Learning (ICML-23)},
  year = {2023},
}

Environment setup

We use conda to manage our environment. To setup the environment, run the command in the terminal:

conda env create -f environment.yml

After that, you can activate it using:

conda activate Fair2Stage

Code file explaination

We exploit two settings in our main experiments:

1. federated reinforcement learning
2. federated online incremental learning

For federated reinforcement learning, the corresponding script is run_2stages_exp_fedrl.py with parameters as follows:

--d: the name of the Atari game, ['Breakout', 'SpaceInvaders', 'Pong']
--b: the temperature parameter used in the softmax
--r: the proportion of nodes being selected each round
--st: the type of score used in the incentive, the default is 'proportion' ['random', 'reverse', 'proportion']
--split: the way to divide the dataset into datasets from different nodes, the default is 'uniform', ['uniform', 'classimbalance', 'powerlaw']
--T: the number of total training iterations
--T1: the number of iterations for contribution evaluation, the default is 0 which uses the hypothesis testing approach to determine the stopping iteration for contribution evaluation. Specify a positive number to stop at a specific iteration
--T2: the number of iterations for giving out incentives, do not need to specify if T is specified
--E: the number of local iterations in each node
--lr: the learning rate
--n: the name of the experiment
--dv: the data valuation method used, the default is cos_grad, ['fed_loo', 'cos_grad', 'mr']
--a: the way to vary the contributions of different nodes, the default is 'normal' which do not vary the contribution of different nodes, ['normal', 'reward_noise', 'state_noise','memory_size','exploration']
--nc: the number of nodes
--alpha: the parameter alpha (i.e., the threshold) for hypothesis testing
--samp_num: the parameter tau (i.e., the number of look-ahead iterations) for hypothesis testing
--samp_dim: the number of nodes used for hypothesis testing

For federated online incremental learning, the corresponding script is run_2stages_exp_stream.py with parameters as follows:

--d: the name of the dataset used, ['mnist', 'cifar10', 'hft', 'electricity', 'pathmnist']
--b: the temperature parameter used in the softmax
--r: the proportion of nodes being selected each round
--st: the type of score used in the incentive, the default is 'proportion' ['random', 'reverse', 'proportion']
--split: the way to divide the dataset into datasets from different nodes, the default is 'uniform', ['uniform', 'classimbalance', 'powerlaw']
--T: the number of total training iterations
--T1: the number of iterations for contribution evaluation, the default is 0 which uses the hypothesis testing approach to determine the stopping iteration for contribution evaluation. Specify a positive number to stop at a specific iteration
--T2: the number of iterations for giving out incentives, do not need to specify if T is specified
--E: the number of local iterations in each node
--lr: the learning rate
--n: the name of the experiment
--dv: the data valuation method used, the default is cos_grad, ['fed_loo', 'cos_grad', 'mr']
--noise: the type of noise used to vary the contributions of the nodes, the default is 'normal' which does not add any noise, ['normal', 'label_noise_different', 'feature_noise_different', 'powerlaw', 'missing_values',nonstatinary_label_noise_inc', 'nonstatinary_label_noise_dec', 'nonstatinary_label_noise_both']
--nc: the number of nodes
--max_noise: the maximum magnitude of the noise added to the dataset
--alpha: the parameter alpha (i.e., the threshold) for hypothesis testing
--samp_num: the parameter tau (i.e., the number of look-ahead iterations) for hypothesis testing
--samp_dim: the number of nodes used for hypothesis testing

Dataset preparation

For MNIST and CIFAR-10, we integrate the downloading of these two datasets in our code. Therefore, no futher effort is needed to run the following code.

For high frequency trading (HFT) dataset, the dataset is avaliable at https://etsin.fairdata.fi/dataset/73eb48d7-4dbc-4a10-a52a-da745b47a649. You can download it and unzip the Benchmark folder to ./datasets.

For electricity loads time series prediction (ELECTRICITY), the dataset is avaliable at https://data.open-power-system-data.org/time_series/2020-10-06. You can directly run the jupyter note book in the ./datasets folder to download and pre-process it.

For PathMNIST(PATH), the homepage of the dataset is at https://medmnist.com/. You can download the pathmnist.npz file from https://zenodo.org/record/6496656#.YoXgofNBz0o and put it under ./datasets.

Examples

For each script, we provide an example to run it.

Federated reinforcement learning

python run_2stages_exp_fedrl.py -d Breakout -T 450 -nc 10 -alpha 0.95 -a state_noise -n state

Federated online incremental learning

python run_2stages_exp_stream.py -d mnist -T 150 -noise label_noise_different -max_noise 0.2 -E 1 -b 150 -alpha 0.7