Table 5.

Best hyperparameters found using the validation set in our experiments on heterophilic datasets.

Dataset	Model	lr	weight decay	depth	hidden	dropout	α	p
Texas	HH-GCN	0.0291	0.0096	2	64	0.8058	0.0043	0.9526
	HH-GraphSAGE	0.0170	0.0053	2	64	0.1967	0.9397	0.7140
	HH-GAT	0.0328	0.0066	2	32	0.1288	0.0902	0.9841
Wisconsin	HH-GCN	0.0105	0.0002	3	128	0.6612	0.9937	0.7140
	HH-GraphSAGE	0.0202	0.0042	3	64	0.3462	0.0100	0.6177
	HH-GAT	0.0539	0.0068	3	16	0.2141	0.0026	0.9797
Actor	HH-GCN	0.0313	0.0087	3	64	0.5511	0.0369	0.5466
	HH-GraphSAGE	0.0133	0.0090	3	32	0.3737	0.0116	0.8368
	HH-GAT	0.0009	0.0001	3	128	0.8708	0.0549	0.9594
Squirrel	HH-GCN	0.0053	0.0001	3	128	0.2455	0.0145	0.8257
	HH-GraphSAGE	0.0296	0.0001	2	128	0.8668	0.9474	0.5198
	HH-GAT	0.0027	0.0001	3	64	0.5131	0.9277	0.1549
Chameleon	HH-GCN	0.0318	0.0057	2	128	0.8040	0.0510	0.9986
	HH-GraphSAGE	0.0225	0.0001	2	32	0.7175	0.9834	0.6226
	HH-GAT	0.0012	0.0008	3	64	0.0439	0.9766	0.9386
Cornell	HH-GCN	0.0505	0.0055	2	32	0.4123	0.0145	0.9660
	HH-GraphSAGE	0.0697	0.0018	2	64	0.0697	0.8807	0.5660
	HH-GAT	0.0572	0.0070	2	64	0.0572	0.0710	0.9979