Table 3:

Testing accuracies (%) of GAT-GC variants (the original one and the ones applied with each of our 4 CPA models) on bioinformatics datasets. We highlight the result of the best performed model per dataset. The highlighted results are significantly higher than those from the corresponding Original model under paired t-test at significance level 5%. The proportion P of multisets that hold the properties in Theorem 1 among all multisets is also reported.

Datasets P(%)	MUTAG 56.9	PROTEINS 29.3	ENZYMES 29.4	NCI1 43.3
Original	84.96 ± 7.65	75.64 ± 3.96	58.08 ± 6.82	80.29 ± 1.89
Additive	89.75 ± 6.39	76.61 ± 3.80	58.90 ± 6.96	81.92 ± 1.89
Scaled	89.65 ± 7.47	76.44 ± 3.77	58.35 ± 6.97	82.18 ± 1.67
f-Additive	90.34 ± 6.05	76.60 ± 3.91	59.80 ± 6.18	81.96 ±2.01
f-Scaled	90.44 ± 6.44	76.81 ± 3.77	58.45 ± 6.35	82.28 ± 1.81