feat: verified gold lineages, held out into dev and test

[hedmana]

Closes #51

What

• fields/gold/gnn.yaml holds 23 papers and 63 originators. Each entry says why it counts.

• The rule: an originator is a work the paper's method is built on, extends, or makes trainable. Excluded are tools, datasets, benchmark splits, baselines, related work, and approaches the paper replaces.

• How entries were checked: against the sentences in which the paper cites the work, taken from Semantic Scholar's citation contexts; where those say too little, against the paper's abstract.

• Examples:

  • GCN keeps ChebNet, Bruna and Hammond ("Our method is based on spectral graph convolutional neural networks, introduced in Bruna et al."), and drops Planetoid ("We choose the same dataset splits as in Yang et al.").
  • GAT keeps Bahdanau ("the particular attentional setup utilized by us closely follows the work of Bahdanau et al.").

• Coverage: 11 GNN-era papers join, including GAT, MPNN, SGC, DGI, R-GCN, DiffPool, FastGCN and VGAE. Six draft papers left:

  • five whose citations Semantic Scholar has no contexts for: Computational Capabilities of GNNs, NN4G, the recursive-versus-graph-NN comparison, GraRep and SDNE;
  • Lusci, whose citations come with nothing to judge by.

  GIN was considered and left out: its central originator, the 1968 WL paper, has no identifier anywhere.

• Splits: each paper is dev (may steer changes to the ranking) or test (only reports it). The split alternates in publication order and was fixed before any tuning.

• gnn.lineages now reports dev and test apart. Alongside recall, it reports gold found as a count over all originators, because recall over reachable originators moves when coverage moves.

Why

The draft was unverified, and a gold set that steers changes can't also judge them. The served ranking against the verified set:

Scorer	dev found	dev Recall@10 / nDCG@10	test found	test Recall@10 / nDCG@10
learned	22/32	0.806 / 0.650	16/31	0.523 / 0.340
equal-weight walk	22/32	0.806 / 0.571	14/31	0.474 / 0.344
time_decayed_pagerank	21/32	0.778 / 0.567	14/31	0.474 / 0.303
gateway	5/32	0.106 / 0.090	4/31	0.167 / 0.080
in_degree	2/32	0.097 / 0.092	2/31	0.076 / 0.038
path_weight	0/32	0 / 0	0/31	0 / 0

Coverage: dev 28/32 and test 30/31 reachable. The learned top 10 overlaps 93 to 95% across seeds.

Verify

uv run pytest tests/test_lineages.py
uv run python -m gnn.lineages --field gnn   # about 4 minutes

Generated with Claude Code