RUNNER

This is the code repository of paper RUNNER: Responsible UNfair NEuron Repair for Enhancing Deep Neural Network Fairness.

This repository implements RUNNER and a variety of baselines on the Adult dataset. More code on other datasets would be released after acceptance.

Operation

For example, run Vanilla train

python main.py --method van --mode dp --lam 1

run RUNNER

python main.py --method NeuronImportance_GapReg --mode eo --lam 5 --neuron_ratio 0.05

Implementations

The hidden size of MLP is 200. We use Adam as the learning optimizer and the batch size is set to 1000 for the $\Delta$ DP metric and 2000 for the $\Delta$ EO metric. The learning rate is set as 0.001. We use Adam as the learning optimizer and the batch size is set as 64 for the $\Delta$ DP metric and 128 for the $\Delta$ EO metric. The learning rate is set as 0.0001. To ensure convergence, on the tabular datasets, for the Vanilla, Oversample, Reweighing, RUNNER, FairSmote, and ROC methods, we train 5 epochs. For the FairNeuron method, we train 10 epochs to ensure convergence. And for the adversarial method, we follow previous work and we train 15 epochs. For the image datasets, we train Vanilla, Oversample, and RUNNER for 10 epochs and train the adversarial method for 20 epochs.

Metric Note

We modified the equalized_odds_difference in Fairlearn by replacing the maximum operation with a summation, in order to align the metric with the one used in the paper.

Other Parameters for RUNNER

For easier comparison, we select hyper-parameters for each method to enable the trained models to have relatively close AP values. For example, to achieve this purpose, the $\lambda$ is set as 1.0 and 0.3 for the $\Delta$ EO and $\Delta$ DP on the Adult dataset. For our method RUNNER, we set the hyper-parameter k as follows:

	Adult	COMPAS	Credit	LSAC	CelebA (wavy)	CelebA (attractive)
DP	50%	5%	5%	5%	20%	5%
EO	5%	5%	5%	5%	50%	20%

Different from Vanilla, Oversample, Reweighing, and FairSmote, other methods rely on hyper-parameters setting. We introduce the hyper-parameter settings as follows:

Other Parameters for Adversarial

The learning rate for the adversary is 1e-4. The training loss is L = $L_{cls}$ + $\lambda$ $L_{adv}$. The $\lambda$ is set as 0.5.

Other Parameters for FairNeuron

We follow FairNeuron to conduct a comparison experiment of these hyperparameters. 𝜃 varies between the interval [1e-4, 1] and 𝛾 varies between the interval [0.5, 1]. Note that we use logarithmic coordinates for 𝜃 since its value is sampled proportionally.

Other Parameters for Decision theory for discrimination-aware classification.

The threshold $\theta$ is set as 0.6.