Co-MLHAN: contrastive learning for multilayer heterogeneous attributed networks

Applied Network Science

Table 5 Results on IMDb-MLH (1000 features), with and without early-stopping

Method	Early-stopping	F1 micro	F1 macro	AUC	F1 ‘action’	F1 ‘comedy’	F1 ‘drama’
Co-MLHAN	No	0.7174 ± 0.0158	0.6406 ± 0.0513	0.8017 ± 0.0225	0.4452 ± 0.1402	0.7375 ± 0.0124	0.7390 ± 0.0133
Co-MLHAN-SA	No	0.6980 ± 0.0143	0.6336 ± 0.0444	0.8004 ± 0.0193	0.4672 ± 0.1233	0.7176 ± 0.0126	0.7159 ± 0.0129
HeCo	No	0.5250 ± 0.0068	0.3595 ± 0.0107	0.5880 ± 0.0112	0.0083 ± 0.02540	0.6118 ± 0.0169	0.4583 ± 0.0190
NSHE	No	0.5778 ± 0.0146	0.4900 ± 0.0198	0.6673 ± 0.0145	0.2587 ± 0.0493	0.5963 ± 0.0168	0.6150 ± 0.0167
Co-MLHAN	Yes	0.7141 ± 0.0159	0.6151 ± 0.0723	0.7975 ± 0.0238	0.3682 ± 0.1987	0.7400 ± 0.0127	0.7370 ± 0.0158
Co-MLHAN-SA	Yes	0.6969 ± 0.0153	0.6089 ± 0.0493	0.7958 ± 0.0204	0.3849 ± 0.1340	0.7242 ± 0.0114	0.7175 ± 0.0142
HeCo	Yes	0.5287 ± 0.0056	0.3626 ± 0.0093	0.5903 ± 0.0097	0.0056 ± 0.0227	0.6150 ± 0.0055	0.4671 ± 0.0081
NSHE	Yes	0.5821 ± 0.0139	0.4886 ± 0.0493	0.6857 ± 0.0108	0.2450 ± 0.0331	0.5954 ± 0.0169	0.6253 ± 0.0164