Table 8:

Results on the test data sets with regard to fairness. 10% of all vertexes are used for training. For fairness metrics ${\Delta }_{SP}$ and ${\Delta }_{EO}$, lower values indicate better performance.

data set	Method	AUROC	F1	ΔSP(↓)	ΔEO(↓)
German Credit	SetGNN	59.16 ± 2.51	81.84 ± 0.93	2.65 ± 5.62	4.06 ± 6.76
	A2	59.81 ± 3.00	82.26 ± 0.13	0.55 ± 0.95	0.78 ± 0.70
	A4	59.66 ± 3.83	80.54 ± 3.52	3.03 ± 6.54	5.07 ± 7.81
	A6	59.88 ± 3.04	82.36 ± 0.38	0.95 ± 0.92	0.47 ± 0.56
Recidivism	SetGNN	96.51 ± 0.48	89.84 ± 0.97	8.63 ± 0.50	4.16 ± 0.51
	A2	96.34 ± 0.39	90.09 ± 0.53	8.53 ± 0.52	3.92 ± 0.68
	A4	96.45 ± 0.35	89.75 ± 0.68	8.49 ± 0.27	3.49 ± 0.66
	A6	96.55 ± 0.54	89.22 ± 0.55	8.51 ± 0.25	3.13 ± 0.64
Credit defaulter	SetGNN	73.46 ± 0.17	87.91 ± 0.27	2.79 ± 0.99	0.98 ± 0.69
	A2	73.43 ± 0.27	87.82 ± 0.24	2.64 ± 1.32	0.93 ± 0.87
	A4	73.58 ± 0.19	87.92 ± 0.25	2.84 ± 1.14	1.38 ± 0.32
	A6	73.78 ± 0.16	88.03 ± 0.14	2.58 ± 0.91	0.81 ± 0.37