Leap of FAITH from GNN-to-MLP: Fairness Aware Inference via disTillation of GrapH Knowledge

Graph Neural Networks (GNNs) exploit graph structure effectively but require costly neighborhood aggregation at inference. GNN-to-MLP knowledge distillation enables graph-free deployment, yet existing methods fail to transfer the heterogeneous smoothing induced by message passing and may propagate the bias captured by the teacher model. We introduce \textbf{FAITH}, a structure- and fairness-aware distillation framework that transfers localized Dirichlet-energy patterns, promotes individual fairness through structured $\ell_{2,1}$ regularization, and extends to group fairness through counterfactual invariance. Across 11 graph benchmarks, FAITH improves structural fidelity and the utility-fairness trade-off while retaining efficient graph-free inference.