Research overview

Statistics

Statistics about the papers and benchmark results in Graph Learning Benchmarks. Publication trends are generated from the currently exported paper set. Research analyses below use frozen experiment snapshots and include their cohort and counting rules.

1150papers with results 2738papers in usage study 84datasets in SOTA analysis

Papers over time

Monthly publication counts for papers with benchmark results in the currently exported site data.

Earliest: Propagation Kernels (Oct 13, 2014)

Latest: Graph Cascades: Contagion-Based Mesoscopic Rewiring for Structure-Aware Graph Machine Learning (Jun 3, 2026)

Benchmark usage in saved extractions

Counts of papers with saved extracted result rows for each benchmark group. The ablation includes some copied, quarantined, and nonstandard rows, so it is an alternative counting lens rather than a corrected total.

Ablation cohort: 2738 dated papers through 2026-06-03. A paper can contribute to several benchmark groups.

How often do reported methods match or beat an MLP?

For each paper and dataset group, the best reported non-MLP result is compared with that paper’s MLP result.

59.3%better or tied 4562better comparisons 1304ties 4026worse comparisons

Dataset results, ordered by worse percentage

Better Tied Worse

Cornell (Heterophilic Graphs (legacy)) 581 comparisons

83.0%

Actor (Heterophilic Graphs (legacy)) 491 comparisons

82.7%

Texas (Heterophilic Graphs (legacy)) 618 comparisons

77.2%

Coauthor-Physics (Classic) 282 comparisons

73.4%

PubMed (Classic) 885 comparisons

72.8%

Wisconsin (Heterophilic Graphs (legacy)) 219 comparisons

61.2%

Amazon-ratings (Heterophily Benchmark (non-legacy)) 212 comparisons

53.8%

CiteSeer (Classic) 901 comparisons

49.8%

Coauthor-CS (Classic) 209 comparisons

43.5%

Pokec (LINKX Benchmarks) 54 comparisons

42.6%

Roman-empire (Heterophily Benchmark (non-legacy)) 234 comparisons

38.9%

Questions (Heterophily Benchmark (non-legacy)) 168 comparisons

36.3%

Aggregated results by dataset group

Heterophilic Graphs (legacy)6 datasets

68.0%

Classic9 datasets

39.6%

Heterophily Benchmark (non-legacy)5 datasets

30.5%

LINKX Benchmarks3 datasets

26.5%

Heterogeneous Graph Benchmarks3 datasets

13.4%

OGB4 datasets

7.1%

LRGB2 datasets

2.9%

GNNBenchmark2 datasets

1.2%

Dataset results show the 12 datasets with the highest worse share. Group percentages pool dataset-level results; papers may appear in multiple datasets and metrics.

GNN vs GT

Stacked dataset counts for SOTA ownership by model architecture. Each snapshot includes datasets with at least 5 papers.

Latest snapshot (2026-06-03)

61datasets with a GNN winner 68datasets with a GT winner

GT combines graph transformers and hybrid architectures. Latest counts include every dataset with that winner type, so categories can overlap. The stacked areas split each dataset’s weight across its tied winning models, matching the paper figure’s fractional ownership counts.

Std reporting

Share of papers with stored results that report a finite, nonzero standard deviation in at least one result.

992 / 1276papers reporting std 77.7%overall reporting rate

Each paper is counted once, in its publication quarter. The denominator is papers with stored results and a parseable publication date; manual standard-deviation overrides are excluded.

Browse Papers Conference Statistics