Multilevel Feature Fusion-Based GCN for Rumor Detection with Topic Relevance Mining

Table 3

Rumor/nonrumor detection results.

Model	Class	Pheme
		Acc.	Prec.	Rec.	F1
SVM-TS	NR	0.685	0.7580	0.7620	0.7570
	R		0.5530	0.5390	0.5390
BURvNN	NR	0.7683	0.7828	0.7436	0.7622
	R		0.7562	0.7930	0.7738
TDRvNN	NR	0.7043	0.6743	0.8177	0.7342
	R		0.7778	0.5909	0.6575
GAN-GRU	NR	0.7513	0.7561	0.7419	0.7494
	R		0.7474	0.7607	0.7538
Bi-GCN	NR	0.8240	0.8610	0.8720	0.8650
	R		0.7530	0.7340	0.7410
AARD	NR	0.8393	0.8259	0.8693	0.8445
	R		0.8652	0.8091	0.8320
AARD-PARG	NR	0.8484	0.8631	0.8290	0.8448
	R		0.8374	0.8677	0.8515
GACL	NR	0.8500	0.8710	0.9010	0.8850
	R		0.8010	0.7500	0.7720
One-level MFF-GCN	NR	0.8547	0.8896	0.8910	0.8899
	R		0.7881	0.7864	0.7859
Two-level MFF-GCN	NR	0.8587	0.8995	0.8845	0.8919
	R		0.7824	0.8078	0.7948
Three-level MFF-GCN	NR	0.8632	0.8989	0.8932	0.8960
	R		0.7951	0.8043	0.7995
One-level MFF-GCN (BERT)	NR	0.8748	0.9147	0.8937	0.9041
	R		0.8024	0.8382	0.8199
Two-level MFF-GCN (BERT)	NR	0.8855	0.9273	0.8968	0.9118
	R		0.8117	0.8635	0.8368
Three-level MFF-GCN (BERT)	NR	0.8973	0.9404	0.9016	0.9206
	R		0.8230	0.8889	0.8547

Note: MFF-GCN (BERT) means pretrained MFF-GCN with BERT. Bold values represent the optimal results for this dataset.