notesum.ai

Published at November 19

Benchmarking Positional Encodings for GNNs and Graph Transformers

cs.AI

Released Date: November 19, 2024

Authors: Florian Grötschla, Jiaqing Xie¹, Roger Wattenhofer¹

Aff.: ¹ETH Zurich

Arxiv: http://arxiv.org/abs/2411.12732v1

Refer to caption

Model	COCO-SP $\uparrow$		PCQM-Contact $\uparrow$		PascalVOC-SP $\uparrow$		Peptides-func $\uparrow$		Peptides-struct $\downarrow$
GCN (Tönshoff et al., 2023)	$13.38\pm 0.07$		$45.26\pm 0.06$		$0.78\pm 0.31$		$68.60\pm 0.50$		$24.60\pm 0.07$
GINE (Tönshoff et al., 2023)	$21.25\pm 0.09$		$46.17\pm 0.05$		$27.18\pm 0.54$		$66.21\pm 0.67$		$24.73\pm 0.17$
GatedGCN (Tönshoff et al., 2023)	$29.22\pm 0.18$		$\mathbf{46.70\pm 0.04}$		$38.80\pm 0.40$		$67.65\pm 0.47$		$24.77\pm 0.09$
CRaWl (Tönshoff et al., 2021)	-		-		$\mathbf{45.88\pm 0.79}$		$\mathbf{70.74\pm 0.32}$		$25.06\pm 0.22$
S²GCN (Geisler et al., 2024)	-		-		-		$\mathbf{73.11\pm 0.66}$		$\mathbf{24.47\pm 0.32}$
DRew (Gutteridge et al., 2023)	-		$34.42\pm 0.06$		$33.14\pm 0.24$		$\mathbf{71.50\pm 0.44}$		$25.36\pm 0.15$
Graph ViT (He et al., 2023)	-		-		-		$68.76\pm 0.59$		$\mathbf{24.55\pm 0.27}$
GatedGCN-VN (Rosenbluth et al., 2024)	$\mathbf{32.44\pm 0.25}$		-		$\mathbf{44.77\pm 1.37}$		$68.23\pm 0.69$		$24.75\pm 0.18$
Exphormer	$\mathbf{34.85\pm 0.11}$	(ESLapPE)	$\mathbf{47.37\pm 0.24}$	(LapPE)	$42.42\pm 0.44$	(LapPE)	$64.24\pm 0.63$	(LapPE)	$24.96\pm 0.13$	(LapPE)
GraphGPS	$\mathbf{38.91\pm 0.33}$	(RWSE)	$\mathbf{46.96\pm 0.17}$	(LapPE)	$\mathbf{45.38\pm 0.83}$	(ESLapPE)	$66.20\pm 0.73$	(LapPE)	$24.97\pm 0.24$	(LapPE)
SparseGRIT	$19.76\pm 0.38$	(noPE)	$45.85\pm 0.11$	(LapPE)	$35.19\pm 0.40$	(GCKN)	$67.02\pm 0.80$	(RRWP)	$24.87\pm 0.14$	(LapPE)
GRIT	$21.28\pm 0.08$	(RWDIFF)	$46.08\pm 0.07$	(SignNet)	$35.56\pm 0.19$	(noPE)	$68.65\pm 0.50$	(RRWP)	$\mathbf{24.54\pm 0.10}$	(RRWP)