notesum.ai

Published at October 22

Can Large Language Models Act as Ensembler for Multi-GNNs?

cs.AI

eess.SP

Released Date: October 22, 2024

Authors: Hanqi Duan¹, Yao Cheng, Jianxiang Yu¹, Xiang Li

Aff.: ¹East China Normal University, Shanghai, China

Arxiv: https://arxiv.org/abs/2410.16822v1

Refer to caption

Model type	Model	Cora (Acc/%)	PubMed (Acc/%)	ogbn-arXiv (Acc/%)
MLP	MLP	66.42	82.40	61.47
GNN	GCN	85.97	83.78	70.81
	GAT	86.71	83.59	70.05
	GIN	85.60	82.03	67.03
	GraphSAGE	84.87	87.01	70.35
	Graphormer	80.41	88.75	72.81
	GT	86.42	88.24	-
	CoarFormer	88.69	89.75	71.66
Language Model	BERT	80.15	93.91	72.78
	SentenceBERT	78.82	92.49	71.42
	DeBERTa	77.79	93.45	72.90
LLM Based	DGTL	81.10	87.10	-
	SNS-GPT4	82.50	93.80	74.40
	InstructGraph-INS	89.33	81.09	-
	InstructGLM	90.77	94.62	75.70
	GAugLLM	-	83.68	74.15
	Baichuan2-13B	13.65	36.04	4.79
	LensGNN-[GCN+GAT]	88.56	94.11	75.78
	LensGNN-[GCN+GIN]	91.88	94.47	75.89
	LensGNN-[GAT+GIN]	88.19	95.43	74.31
	LensGNN-ALL	90.40	95.68	75.91