Kipf and Welling, University of Amsterdam (ICLR 2017) — ARXIV-1609.02907
Key takeaways
- The layer: multiply node features by the degree-normalized adjacency with self-loops (the “renormalization trick”), then by a weight matrix. Each layer mixes one hop of neighborhood, at cost linear in edges
- Propagation ablation (Table 3, mean of 100 initializations on the Planetoid split): the renormalized form is best on all three citation graphs. Chebyshev K = 2 reaches 81.2 on Cora and 73.8 on Pubmed, against 81.5 and 79.0
- Results (Table 2): 70.3 on Citeseer, 81.5 on Cora and 79.0 on Pubmed, against Planetoid’s 64.7, 75.7 and 77.2
- Depth (appendix B): accuracy peaks at 2–3 layers. Beyond about 7, training without residual connections becomes difficult
Standing in the anthology
Filed from #163 (“graph representation / gnn”) as the reference
architecture the other GNN papers here measure against. GIN
(LIT-580) and Shchur et al. (LIT-581) are compared_against it.
No practice from it. Its numbers come from one fixed split, which Shchur et al. show can reorder models, and later work under fair tuning finds it a hard baseline to beat. That finding is filed from the papers that measured it.
Skimmed: the layer, the propagation ablation and the depth appendix. No NOTE is filed.