The final output layer is usually linear, which rigidly limits the reachable solutions (here illustrated on distributions).

See paper for details. Necessary embedding dimensions quickly become unrealistic for large graphs.

Equipping models with this layer improves their performance on large and dense datasets.