Figure 2
The scatter plot is drawn on a coordinate plane. The horizontal axis is labeled “Simulated centrality” and ranges from 0.0 to 1.0 in increments of 0.2 units. The vertical axis is labeled “Absolute prediction error” and ranges from 0.0 to 1.0 in increments of 0.2 units. A legend titled “Country” identifies two groups: “Large” shown with red cross markers, and “Small” shown with blue cross markers. Data points are distributed across the full range of simulated centrality from 0.0 to 1.0. For both groups, the majority of points cluster in the lower region of the plot, with absolute prediction error values between 0.0 and 0.3. A moderate density of points appears between 0.3 and 0.6, while fewer points extend into higher error values above 0.6, reaching up to 1.1. The “Large” country points are more widely dispersed vertically, including several higher error observations above 0.8, while the “Small” country points are more concentrated below 0.6 with fewer extreme values. Note: All numerical data values are approximated.Prediction error vs. centrality. Note(s): This scatterplot shows how the GNN prediction error is negatively correlated with node centrality. Agents in more central network positions exhibit lower prediction errors, demonstrating the model’s ability to extract relational signals effectively. Source(s)