A Generalization of Transformer Networks to Graphs

AAAI Workshop on DLG, 2021
← Back to References

Bibliographic Information

TitleA Generalization of Transformer Networks to Graphs
AuthorsVijay Prakash Dwivedi, Xavier Bresson
VenueAAAI Workshop on DLG
Year2021
Linkhttps://arxiv.org/abs/2012.09699

Summary

Introduces Laplacian Positional Encoding (LapPE) for graph transformers using eigenvectors of the graph Laplacian. Suffers from sign ambiguity (later addressed by SignNet). Requires symmetric matrices, making it inapplicable to directed graphs — a limitation STRATA's reachability profiles do not have.