Geometrical fairness tweaks Laplacian to reduce bias in graph neural networks

Graph-based learning methods, including graph neural networks (GNNs), have become essential in fields like social networks, recommendation systems, and drug discovery. However, they risk propagating or amplifying biases present in training data, raising fairness concerns. In a new arXiv paper, Pérez-Peralta et al. tackle this by introducing a fairness-aware adaptation of graph-based diffusion. Their approach modifies the underlying Laplacian operator—a key mathematical component of graph models—through multiple complementary transformations: subspace projections, spectral adjustments, and frequency-based filtering. These techniques specifically target and suppress bias-related components in the model, leveraging the natural smoothing properties of graph diffusion. The authors provide a principled theoretical analysis of the resulting fairness properties, grounding their approach in established diffusion dynamics.

Empirically, the method is evaluated on both synthetic benchmarks and real-world graph datasets, demonstrating that it achieves competitive performance on standard graph learning tasks while significantly improving fairness metrics (e.g., demographic parity, equal opportunity). Importantly, these gains come with limited additional computational cost, making the approach practical for deployment. The paper includes 21 tables and 6 figures across 32 pages, with code expected to be released. As GNNs are increasingly used in high-stakes applications like hiring, lending, and policing, this work offers a concrete, low-overway-to embed fairness directly into the model architecture rather than as a post-hoc fix. The method is model-agnostic and can be integrated into existing graph diffusion frameworks, potentially setting a new standard for responsible GNN design.