Note(s): The table compares redistributive and informational performance across three tax design approaches. The linear proxy-based rule reflects conventional targeting mechanisms based on observable characteristics. The XGBoost benchmark uses the same observable attributes as the proposed model but excludes relational information, thereby serving as a strong non-graph machine learning baseline. The GNN-based model incorporates economic network structure through graph embeddings. Welfare losses are computed under distributional ambiguity using worst-case income realizations consistent with inferred information sets. Results are averaged across 1,000 simulation replications. Lower MAE, lower post-tax Gini coefficients, higher Kakwani indices and smaller welfare losses indicate superior performance