Core Concepts
While incorporating unique node identifiers (UIDs) into Graph Neural Networks (GNNs) can theoretically enhance their expressiveness, effectively leveraging these UIDs in practice requires careful regularization to prevent overfitting and promote invariance to the specific UID values.
Stats
SIRI improved the total accuracy on the BREC dataset by 12.3%.
SIRI outperformed all other evaluated methods in 2 out of 4 graph groups on the BREC dataset.
The largest improvement achieved by SIRI on the BREC dataset was in the Regular graphs group.
On the isInTriangle task, SIRI achieved 88.45 ± 2.04% accuracy in the interpolation setting and 78.20 ± 2.53% in the extrapolation setting.
RNI achieved nearly the same accuracy as the baseline on the isInTriangle task (using a constant non-unique identifier), which is unable to solve the task.