Reproduce “A Theory for Compressibility of Graph Transformers for Transductive Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hamed Shirzad, Honghao Lin, Ameya Velingker, Balaji Venkatachalam, David Woodruff, Danica Sutherland
The authors analyze width reduction in Graph Transformers for node classification. They project query and key vectors into a lower-dimensional space to approximately preserve pairwise similarities, which controls attention scores and node-output error under bounded norms and constant depth. On Amazon Photo, the best width-4 model scored 0.944, versus a width-64 mean of 0.953; these were independently trained, not compression.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.