Mean Absolute Error (MAE) results grouped by the connection (Call and SMS) and evaluation types.

A lower MAE value indicates better performance, and the best results are shown in bold. MAE is decomposed to mean (STD) prediction error of positive edges (Positive), mean (STD) prediction error of random negative edges (R-Negative), mean (STD) prediction error of historical negative edges (H-Negative), and average of the three means (Ave). Using Wilcoxon signed-rank test revealed that all the GNN edge prediction results are statistically significantly different compared to the rEdgeBank baseline.