Graph Attention Networks
2018 · paper · AI and machine learning
Introduces graph attention networks, neural networks for graph-structured data in which nodes attend over their neighbors' features with self-attention. The authors report state-of-the-art or matching results on three citation network datasets and a protein-protein interaction dataset.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
0 findings
None that passed a second check.
5 minor (typos and the like; most checked by the judge only)
Typo Page 1 names the same institution two ways: affiliations read "Montréal Institute for Learning Algorithms" while the footnote reads "Montréal Institute of Learning Algorithms".
- p. 1 Author affiliations (Romero, Bengio)
Montréal Institute for Learning Algorithms
- p. 1 Footnote *
Montréal Institute of Learning Algorithms
Judge's note: Viewed page 1 image. Both forms are printed: the Romero and Bengio affiliations use 'for', the asterisk footnote uses 'of'. This is an internal inconsistency on one page; the affiliation form is used twice, so the footnote is the likely slip.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 1 (right) labels the attention coefficients with vector arrows ("⃗α12" etc.), whereas the text (Eq. 2, "αij = softmaxj(eij)") and the left panel write the scalar αij without an arrow.
- p. 4 Figure 1, right panel edge labels
⃗α12
- p. 3 Eq. 2
αij = softmaxj(eij)
Judge's note: Viewed page 4 image. The right-panel edge labels α11 to α16 carry arrows; the left panel and Eqs. 2-6 use α_ij without one. The arrow may be meant to stack the K = 3 per-head coefficients, but neither caption nor text says so. Minor notation inconsistency where either form could be intended.
Second check: none (judge only)
Report a problem with this finding
Typo Text says "The average number of nodes per graph is 2372." while Table 1 gives "56944 (24 graphs)", i.e. 2372.67, which rounds to 2373.
- p. 6 Section 3.1, Inductive learning
The average number of nodes per graph is 2372.
- p. 6 Table 1, PPI column, # Nodes
56944 (24 graphs)
Judge's note: Viewed page 6 image. Table 1 gives 56944 nodes over 24 graphs, and the split 44906 + 6514 + 5524 sums to 56944. The mean is 2372.67, so the printed 2372 is a truncation rather than a rounding. The one-node difference affects nothing else.
Second check: none (judge only)
Report a problem with this finding
Typo Reference Gori et al. (2005) prints the page range without a dash: "pp. 729734, 2005".
- p. 11 References, Gori et al. (2005)
pp. 729734, 2005
Judge's note: Viewed page 11 image. The entry reads 'pp. 729734, 2005' with no dash, so this is not an extraction artefact. Other entries on the page use dashed ranges (e.g. 'pp. 249–256', 'pp. 770–778'). The evident fix is 729–734.
Second check: none (judge only)
Report a problem with this finding
Typo The reference list spells one author's surname two ways: "Ruslan Salakhudinov" (Yang et al., 2016) and "Ruslan Salakhutdinov" (Srivastava et al., 2014).
- p. 12 References, Yang et al. (2016)
Ruslan Salakhudinov
- p. 12 References, Srivastava et al. (2014)
Ruslan Salakhutdinov
Judge's note: Viewed page 12 image. Both spellings are printed, with the same given name and a one-letter difference. This is an internal inconsistency within the reference list; the Yang et al. entry is missing the 't' relative to the Srivastava et al. entry.
Second check: none (judge only)