Why is it that only GCN achieves about 84.9% of the results, but only 84% when attention is added?

qitianwu / SGFormer

The official implementation for NeurIPS2023 paper "SGFormer: Simplifying and Empowering Transformers for Large-Graph Representations"

213 stars 18 forks source link

Why is it that only GCN achieves about 84.9% of the results, but only 84% when attention is added? #14

Closed CharliedoD closed 4 months ago

CharliedoD commented 5 months ago

1715856303417

qitianwu commented 4 months ago

Hi, we use the Final Test (testing score achieved by the epoch with the highest valid score) for all the reported results in the paper. The Highest Test is not used for comparison since it requires testing data