Migrate Demographic Group For Fair GNNs

TL;DR

A brand new framework, FairMigration, is proposed, which can dynamically migrate the demographic groups instead of keeping that fixed with raw sensitive attributes, which balances model performance and fairness well.

Abstract

Graph Neural networks (GNNs) have been applied in many scenarios due to the superior performance of graph learning. However, fairness is always ignored when designing GNNs. As a consequence, biased information in training data can easily affect vanilla GNNs, causing biased results toward particular demographic groups (divided by sensitive attributes, such as race and age). There have been efforts to address the fairness issue. However, existing fair techniques generally divide the demographic groups by raw sensitive attributes and assume that are fixed. The biased information correlated with raw sensitive attributes will run through the training process regardless of the implemented fair techniques. It is urgent to resolve this problem for training fair GNNs. To tackle this problem, we propose a brand new framework, FairMigration, which can dynamically migrate the demographic groups instead of keeping that fixed with raw sensitive attributes. FairMigration is composed of two training stages. In the first stage, the GNNs are initially optimized by personalized self-supervised learning, and the demographic groups are adjusted dynamically. In the second stage, the new demographic groups are frozen and supervised learning is carried out under the constraints of new demographic groups and adversarial training. Extensive experiments reveal that FairMigration balances model performance and fairness well.

Figures (8)

• Figure 1: Visualization of attributes correlation. The cell in the i-th row and j-th column represents the Pearson Correlation Coefficient (PCC) among the i-th attribute and the j-th attribute. The sensitive attribute is marked by the red lines.
• Figure 2: A toy diagram of group migration. $\Delta EO$ is a fairness metric. The lower $\Delta EO$ indicates the more fair performance.
• Figure 3: The change curve of group similarity distribution of vanilla GCN on three datasets. The bold lines with the same color indicate the mean / std of two group similarity distributions on the same dataset, respectively. The shaded areas indicate the group similarity distribution difference. The wider the shaded area on the y-axis, the larger the similarity difference between groups.
• Figure 4: Framework of FairMigration. The workflow of the Self-Supervised Learning (SSL) stage only is colored black. The workflow of the Supervised Learning (SL) stage only is colored red. The workflow of both the SSL stage and SL stage is colored blue. At the SSL stage, FairMigration optimizes the encoder by personalized self-supervised learning and migrates the demographic groups. At the SL stage, FairMigration optimizes the encoder and the classifier under the constraints of migrated groups and adversarial training.
• Figure 5: Flipping the sensitive attribute (age) on credit when testing. The vanilla denotes testing the vanilla GNNs without sensitive attribute flipping. The vanilla-R denotes testing the vanilla GNNs with sensitive attribute flipping. The FM denotes testing FairMigration without sensitive attribute flipping. The FM-R denotes testing FairMigration with sensitive attribute flipping.

Theorems & Definitions (2)

• Definition 1
• Definition 2