Fig. 1.

In the problem of link prediction, we are asked to identify which unobserved links in a network are more likely to exist. Nodes could represent individuals or drugs, and links could represent, respectively, friendship relationships in a social network or harmful drug–drug interactions. In this example, links $AB$ and $CD$ exist but have not been observed, so we aim to predict them. Model 1 pays attention only to the connectivity of nodes, and it captures that many nodes are connected to $A$, so it correctly predicts the $AB$ link. However, since there is nothing especial about the connectivity of $C$ and $D$, it misses the $CD$ link. Conversely, model 2 pays attention only to group structure, so it realizes that all nodes in the group at Right are connected to each other, and it predicts the $CD$ link. However, since in the group at Left many pairs of nodes are not connected to each other, it misses link $AB$. Model stacking, in which models 1 and 2 are combined, may be able to predict both links.