GNN predicted masses, ${\text{M}}_{500}^{\text{GNN}}$, vs true cluster masses, ${\text{M}}_{500}^{\text{True}}$, for the training (left) and testing (right) samples, drawn from the original unaugmented dataset across different redshift bins. The black line shows the self-similar relation. The two subpanels show the fractional residuals; the solid blue line indicates the median and the shaded region represents the 1σ scatter. As expected, the model predictions for the training set are effectively unbiased, while the testing sample exhibits a mild average bias of (1 − b) = 0.96, reflecting redshift-dependent effects.