notesum.ai

Published at November 5

DA-MoE: Addressing Depth-Sensitivity in Graph-Level Analysis through Mixture of Experts

cs.LG

cs.AI

Released Date: November 5, 2024

Authors: Zelin Yao¹, Chuang Liu¹, Xianke Meng¹, Yibing Zhan², Jia Wu³, Shirui Pan⁴, Wenbin Hu¹

Aff.: ¹School of Computer Science, Wuhan University; ²JD Explore Academy; ³School of Computing, Macquarie University; ⁴School of Information and Communication Technology, Griffith University

Arxiv: http://arxiv.org/abs/2411.03025v1

	Biochemical Domain			Social Domain
	PROTEINS	NCI1	MUTAG	IMDB-B	IMDB-M	COLLAB	RDT-B
GIN	$79.91_{\pm 0.62}$	$83.28_{\pm 0.33}$	$95.39_{\pm 0.75}$	$77.77_{\pm 0.59}$	$53.63_{\pm 0.36}$	$82.59_{\pm 0.21}$	$84.24_{\pm 1.75}$
+ DA-MoE	$\textbf{80.27}_{\pm\textbf{0.39}}$	$\textbf{83.57}_{\pm\textbf{0.27}}$	$\textbf{95.93}_{\pm\textbf{0.57}}$	$\textbf{78.78}_{\pm\textbf{0.50}}$	$\textbf{55.09}_{\pm\textbf{0.20}}$	$\textbf{83.89}_{\pm\textbf{0.19}}$	$\textbf{84.72}_{\pm\textbf{1.79}}$
Improvement	$0.45\%$	$0.35\%$	$0.57\%$	$1.30\%$	$2.72\%$	$1.57\%$	$0.57\%$
GCN	$78.68_{\pm 0.37}$	$75.94_{\pm 0.25}$	$87.50_{\pm 0.57}$	$\textbf{80.13}_{\pm\textbf{0.31}}$	$55.25_{\pm 0.32}$	$84.18_{\pm 0.19}$	$86.40_{\pm 1.45}$
+ DA-MoE	$\textbf{78.99}_{\pm\textbf{0.39}}$	$\textbf{80.63}_{\pm\textbf{0.44}}$	$\textbf{93.00}_{\pm\textbf{1.03}}$	$79.92_{\pm 0.39}$	$\textbf{55.95}_{\pm\textbf{0.18}}$	$\textbf{84.62}_{\pm\textbf{0.38}}$	$\textbf{87.07}_{\pm\textbf{0.86}}$
Improvement	$0.39\%$	$6.18\%$	$6.29\%$	–	$1.27\%$	$0.52\%$	$0.78\%$
GatedGCN	$79.22_{\pm 0.25}$	$\textbf{80.24}_{\pm\textbf{0.39}}$	$92.11_{\pm 1.05}$	$78.02_{\pm 0.61}$	$53.33_{\pm 0.40}$	$76.23_{\pm 0.96}$	$81.69_{\pm 0.99}$
+ DA-MoE	$\textbf{79.77}_{\pm\textbf{0.43}}$	$78.40_{\pm 0.48}$	$\textbf{93.56}_{\pm\textbf{0.90}}$	$\textbf{78.13}_{\pm\textbf{0.46}}$	$\textbf{54.78}_{\pm\textbf{0.37}}$	$\textbf{83.25}_{\pm\textbf{0.10}}$	$\textbf{82.23}_{\pm\textbf{1.27}}$
Improvement	$0.69\%$	–	$1.57\%$	$0.14\%$	$2.72\%$	$9.21\%$	$0.66\%$