Structural and positional ensembled encoding for Graph Transformer

Jeyoon Yeom, Taero Kim, Rakwoo Chang, Kyungwoo Song

Abstract: Highlights•We combined information of graph structures to create a novel positional encoding.•The components of the new unified encoding compensate for each other’s shortcomings.•Our new architecture continuously injects unified information into the attention.•Our model performs better when edge information is scarce.•The model discriminates graphs that are structurally indistinguishable.