| EU | US | Diff (%) |
---|
(a) Only text |
EU | 0.775 | 0.771 | 0.5 |
US | 0.711 | 0.745 | 4.7 |
(b) Text + user history |
EU | 0.836 | 0.826 | 1.2 |
US | 0.783 | 0.811 | 3.6 |
(c) Text + network stats |
EU | 0.842 | 0.780 | 7.9 |
US | 0.717 | 0.769 | 7.2 |
(d) Text + raw network |
EU | 0.885 | 0.874 | 1.2 |
US | 0.803 | 0.846 | 5.3 |
(e) All |
EU | 0.888 | 0.876 | 1.4 |
US | 0.832 | 0.847 | 1.8 |
- In each table marked by letters (a–e), columns and rows represent the training and test regions, respectively. Here, raw network represents Walklets (Perozzi et al. 2016), the best-performing node embedding model (see Fig. 11). The diff column shows the AUC difference between training and testing on the same domain versus out-of-domain