Multiview Embedding with Partial Labels to Recognize Users of Devices Based on Unified Transformer

Table 6

Ablation Studies of MVEPL without/with label encoder.

Graph	Transformer	Evaluation
		Accuracy	Precision	Recall	F1	Cost time
rDNS	Transformer	0.7584/0.8362	0.2125/0.3143	0.6124/0.6789	0.3155/0.4297	≈6
	Cotransformer	0.7652/0.8576	0.2200/0.3508	0.6217/0.6656	0.3250/0.4594	≈10
	Unified transformer	0.7921/0.8471	0.2440/0.3420	0.6131/0.7381	0.3491/0.4674	≈13
AS/subnet	Transformer	0.7702/0.8791	0.2268/0.4089	0.6345/0.7401	0.3342/0.5268	≈9
	Cotransformer	0.8227/0.8812	0.2844/0.4123	0.6266/0.7216	0.3912/0.5248	≈13
	Unified transformer	0.8387/0.8954	0.3088/0.4536	0.6254/0.7359	0.4135/0.5613	≈16
Location	Transformer	0.7498/0.8402	0.2006/0.3235	0.5872/0.6942	0.2990/0.4413	≈10
	Cotransformer	0.7302/0.8675	0.1799/0.3777	0.5531/0.7062	0.2715/0.4922	≈12
	Unified transformer	0.8129/0.8677	0.2661/0.3792	0.6019/0.7147	0.3690/0.4955	≈18
Hardware/software	Transformer	0.7977/0.8320	0.2569/0.3031	0.6475/0.6529	0.3679/0.4140	≈5
	Cotransformer	0.7851/0.8452	0.2409/0.3371	0.6338/0.7272	0.3491/0.4607	≈8
	Unified transformer	0.8045/0.8523	0.2652/0.3511	0.6495/0.7362	0.3766/0.4755	≈10
Total	Transformer	0.8245/0.8752	0.2950/0.3992	0.6694/0.7379	0.4095/0.5181	≈25
	Cotransformer	0.8325/0.8825	0.3048/0.4153	0.6575/0.7175	0.4165/0.5261	≈30
	Unified transformer	0.8543/0.9158	0.3491/0.5265	0.6973/0.7337	0.4653/0.6131	≈50

Bold values highlight the best models outperforming in accuracy and F1.