Graph Transformers Show Distance-Misaligned Training in Node Classification
A new study posted on arXiv (2604.22413) delves into the issues that Graph Transformers face regarding their failure mechanisms. While these models excel at globally integrating information, they sometimes mismanage communication based on the distance within the graph. The researchers used a synthetic benchmark for node classification on contextual stochastic block model graphs to examine how training may become misaligned with distance. They found that the preferred bias in graph distance changes depending on the task's locality. Interestingly, an oracle adaptive controller nearly reaches the ideal fixed bias, whereas a task-agnostic zero-gap controller doesn’t perform as well.
Key facts
- arXiv paper 2604.22413 studies Graph Transformers.
- Focuses on distance-misaligned training in node classification.
- Uses synthetic contextual stochastic block model graphs.
- Labels generated by controllable mixture of local and far-shell signals.
- Three main findings reported.
- Preferred graph-distance bias changes with task locality.
- Oracle adaptive controller nearly matches best fixed bias.
- Task-agnostic zero-gap controller is weaker.
Entities
Institutions
- arXiv