Graph Neural Networks learn trivial mini-batch dependent heuristics for link prediction rather than generalizable graph representations

Research shows popular link prediction models can exploit batch-normalisation layers to solve edge classification tasks using mini-batch dependent heuristics rather than learning consistent representations across graphs.

When this bias is corrected, the network representation shows increased alignment with node-class relevant features, suggesting the model learns a graph representation that better matches the underlying graph's properties.

The findings suggest standard link prediction training may overestimate how well link predictors learn generalized graph representations that remain consistent across different tasks.