Table 1.

Increase in supervised performance when using Half-Hop.

	Am. Comp.	Am. Photos	Co.CS	WikiCS
GCN	90.22 ± 0.60	93.59 ± 0.42	94.06 ± 0.16	81.93 ± 0.42
HH-GCN	90.92 ± 0.35	94.52 ± 0.22	94.71 ± 0.16	82.57 ± 0.36
Δ	+0.70 (↑)	+0.93 (↑)	+0.65 (↑)	+0.64 (↑)
GraphSAGE	84.79 ± 1.08	95.03 ± 0.33	95.11 ± 0.10	83.67 ± 0.45
HH-GraphSAGE	86.60 ± 0.49	94.55 ± 0.41	95.13 ± 0.21	82.81 ± 0.32
Δ	+1.81 (↑)	−0.48 (↓)	+0.02 (↑)	−0.86 (↓)

The average and standard deviation of accuracy is computed over 20 random splits and model initializations. The absolute improvement $\left(\mathrm{\Delta }\right)$ is also reported.