Node Importance Estimation with Multiview Contrastive Representation Learning
Table 4
The results of MCRL and its variants.
Data
Metric
MCRL-none
MCRL-score
MCRL
FB15K
NDCG@100
0.9485 ± 0.034
0.9502 ± 0.015
0.9626 ± 0.013
SPEARMAN
0.7901 ± 0.026
0.8173 ± 0.012
0.8229 ± 0.014
HR@100
0.4829 ± 0.094
0.4902 ± 0.075
0.5100 ± 0.069
TMDB5K
NDCG@100
0.9054 ± 0.048
0.9107 ± 0.021
0.9195 ± 0.020
SPEARMAN
0.7768 ± 0.022
0.7826 ± 0.016
0.7961 ± 0.013
HR@100
0.5427 ± 0.069
0.5579 ± 0.038
0.5680 ± 0.041
IMDB
NDCG@100
0.9422 ± 0.016
0.9535 ± 0.007
0.9624 ± 0.008
SPEARMAN
0.7593 ± 0.014
0.7664 ± 0.009
0.7713 ± 0.007
HR@100
0.5545 ± 0.067
0.5636 ± 0.038
0.5720 ± 0.044
MCRL-none: the model without score aggregation module. MCRL-score: the model that simply splices the scores of nodes with the features of edges to calculate attention weights. MCRL: our proposed model, which splices the features of nodes with those of edges to calculate attention weights.
