Adversarial Graph Neural Networks for Fair Insurance Pricing: An Integrative Framework with a Synthetic Benchmark
Narendra Singh Chilwal · NORMA · 2025
The increasing complexity of auto-insurance markets is pushing carriers toward machine-learning pricing models that estimate individual risk from rich, relational data. Graph Neural Networks (GNNs) excel at capturing structures such as geographic proximity or shared accident involvement, yet, when trained on historical claims—they can inherit the very inequities regulators now scrutinize. We present an end-to-end framework for fair premium estimation that combines (i) a relational GNN for claim-frequency prediction with (ii) in-processing adversarial debiasing across multiple protected attributes. To evaluate trade-offs systematically, we generate a 50,000-node Bias-on-Demand benchmark whose representation, measurement, and structural biases are all user-controlled. Our architecture inserts Gradient-Reversal Layers behind the GNN encoder and attaches adversary heads for race, gender, age, and geographic quadrant. A single hyper-parameter λ balances predictive loss against the four adversarial losses, tracing the fairness–accuracy frontier in one sweep. On the synthetic benchmark, the untuned GNN attains RMSE = 2.62 but shows marked disparities (DPrace = 1.33, DPgeo = 1.65). With λ = 1, single-attribute adversaries reduce their respective gaps by ∼25–35% with ≤1% RMSE increase. A unified multi-attribute adversary at the same λ halves all four DP metrics (with an ∼9% RMSE increase), and an intersectional race×gender head at λ = 2 drives DPrace×gender to 0.24 (with an ∼10% RMSE increase). A resource-constrained real-data prototype (15k-node classification on Insurance.arff, city-based graph) shows the same disparity ranking (geography dominant) with DPCity ≈ 0.75 and AUC ≈ 0.49 (Accuracy ≈ 0.62), indicating qualitative transfer but also highlighting higher absolute bias; reported DP is unpruned (upper bound) due to many small city cells. These results demonstrate that adversarial in-processing inside GNNs – together with a reproducible, bias-parameterised benchmark – offers a practical route to equitable, regulator-ready insurance pricing. Future work will integrate post-processing calibration to close the residual geo gap and scale training to full portfolios via mini-batch GNNs.