Learning Fair Graph Representations with Multi-view Information Bottleneck

TL;DR

FairMIB addresses fairness in graph neural networks by decomposing graphs into feature, structural, and diffusion views and optimizing a multi-view information bottleneck to balance utility and fairness. It leverages cross-view mutual information, a multi-view conditional information bottleneck, and an IPW-based diffusion correction to block bias propagation, reinforced by a contrastive multi-view consistency constraint. The approach achieves state-of-the-art fairness-utility performance on five real-world datasets, validated through extensive ablations and sensitivity analyses. This framework enables robust, bias-mitigated graph representations suitable for high-stakes tasks in domains with sensitive attributes.

Abstract

Graph neural networks (GNNs) excel on relational data by passing messages over node features and structure, but they can amplify training data biases, propagating discriminatory attributes and structural imbalances into unfair outcomes. Many fairness methods treat bias as a single source, ignoring distinct attribute and structure effects and leading to suboptimal fairness and utility trade-offs. To overcome this challenge, we propose FairMIB, a multi-view information bottleneck framework designed to decompose graphs into feature, structural, and diffusion views for mitigating complexity biases in GNNs. Especially, the proposed FairMIB employs contrastive learning to maximize cross-view mutual information for bias-free representation learning. It further integrates multi-perspective conditional information bottleneck objectives to balance task utility and fairness by minimizing mutual information with sensitive attributes. Additionally, FairMIB introduces an inverse probability-weighted (IPW) adjacency correction in the diffusion view, which reduces the spread of bias propagation during message passing. Experiments on five real-world benchmark datasets demonstrate that FairMIB achieves state-of-the-art performance across both utility and fairness metrics.

Figures (5)

• Figure 1: Overview of the proposed FairMIB framework. The model first disentangles the input graph into three complementary views: a Diffusion View, a Feature View, and a Structural View. Each view is encoded by a dedicated variational encoder to obtain a latent representation. These representations are then fused through a Projector, producing a fair representation that is concatenated with the sensitive attribute $\mathbf{S}$ during training to guide the Decoder toward fair predictions.
• Figure 2: Ablation study and Multi-view study about FairMIB
• Figure 3: Parameter sensitivity results on Pokec datasets. Results demonstrate that FairMIB achieves stable performance across a wide range of parameter settings.
• Figure 4: The generation process for the Diffusion View begins with the original attributed graph. First, attribute bias is balanced using Inverse Propensity Weighting (IPW). A K-step fairness diffusion process is then executed on this basis to ultimately generate the Diffusion View, which incorporates fair neighborhood information.
• Figure 5: Parameter sensitivity results on two datasets.

Theorems & Definitions (2)

• Definition 1: View Redundancy
• Corollary 1: Representation Sufficiency