Table 4.

Increase in performance when using Half-Hop with different SSL frameworks.

	input (test)	encoder	Am. Comp.	Am. Photos	Co.CS	Co.Phy	WikiCS
GRACE	G	2-GCN	89.53 ± 0.35	92.78 ± 0.45	91.12 ± 0.20	OOM	80.14 ± 0.48
HH-GRACE	G	2-GCN	91.11 ± 0.18	94.21 ± 0.26	93.59 ± 0.16	OOM	79.77 ± 0.40
HH-GRACE	HH(G)	2-GCN	90.65 ± 0.19	94.89 ± 0.23	94.76 ± 0.14	OOM	80.15 ± 0.16
BGRL	G	2-GCN	90.34 ± 0.19	93.17 ± 0.30	93.31 ± 0.13	95.73 ± 0.05	79.98 ± 0.10
HH-BGRL	G	2-GCN	91.02 ± 0.27	93.88 ± 0.19	93.61 ± 0.13	95.75 ± 0.13	80.76 ± 0.71
HH-BGRL	HH(G)	2-GCN	90.94 ± 0.19	94.50 ± 0.35	94.74 ± 0.15	96.13 ± 0.10	80.37 ± 0.62
BGRL	G	3-GCN	90.04 ± 0.23	92.59 ± 0.34	92.42 ± 0.17	95.32 ± 0.51	78.22 ± 0.77
HH-BGRL	G	3-GCN	90.53 ± 0.27	93.09 ± 0.16	92.58 ± 0.20	95.45 ± 0.09	79.76 ± 0.61
HH-BGRL	HH(G)	3-GCN	91.10 ± 0.21	94.34 ± 0.25	94.76 ± 0.12	96.10 ± 0.09	81.11 ± 0.48

Encoders 2-GCN and 3-GCN represent a 2 layer and a 3 layer GCN respectively. Input (test) denotes the graph supplied at test time, where we can choose to use the original graph G or the augmented graph $\mathrm{H}\mathrm{H}\left(\mathrm{G}\right)$. Performance is reported in terms of classification accuracy along with standard deviations. All experiments are performed over 20 random dataset splits and model initializations.