notesum.ai

Published at December 10

A Dynamical Systems-Inspired Pruning Strategy for Addressing Oversmoothing in Graph Neural Networks

cs.LG

cs.AI

Released Date: December 10, 2024

Authors: Biswadeep Chakraborty, Harshit Kumar, Saibal Mukhopadhyay

Arxiv: http://arxiv.org/pdf/2412.07243v1

Refer to caption

	Metric	Cora	Citeseer	Cornell
	Nodes $(N)$	2,708	3,327	183
Models	Edges $(E)$	5,429	4,732	280
	Avg. Degree $(2\|E\|/\|N\|)$	4.01	2.84	3.06
GCN	Best Accuracy	81.5	75.7	54.2
	#Layers	2	2	2
	GFLOPS	0.598	1.789	0.049
	Accuracy/GFLOPS	136.28	43.53	1106.12
GAT	Best Accuracy	82.55	76.1	56.3
	#Layers	4	2	2
	GFLOPS	2.351	6.754	0.184
	Accuracy/GFLOPS	35.11	11.27	306.52
G2GAT	Best Accuracy	83.27	82.06	61.55
	#Layers	128	128	128
	GFLOPS	1.209	2.452	0.0879
	Accuracy/GFLOPS	68.88	33.47	700.34
DYNAMO-GAT	Best Accuracy	83.21	82.01	62.56
	#Layers	128	128	128
	GFLOPS	0.605	1.675	0.051
	Accuracy/GFLOPS	137.53	48.96	1226.67