NodeGround: A Node Classification Benchmark in the Graph Foundation Model Era

Jinmo Lee; Dooho Lee; Minho Jeong; Jaemin Yoo

cs.LGSep 30, 2026

NodeGround: A Node Classification Benchmark in the Graph Foundation Model Era

Authors: Jinmo Lee, Dooho Lee, Minho Jeong, Jaemin Yoo

Organizations: Nums AI

Abstract

Can a pretrained graph model replace training and tuning a separate predictor for each dataset? Answering this requires evaluating prediction quality alongside computational cost. We present NodeGround, a node classification benchmark that puts graph foundation models (GFMs) and dataset-specific supervised learning under a common evaluation framework. The benchmark spans 51 datasets and evaluates six GFMs alongside 15 supervised methods under two label-availability regimes. Shared data partitions, validation-only model selection, controlled hyperparameter searches, and multiple predictive metrics make comparisons systematic, while workflow measurements account for adaptation, training, tuning, and inference. The results favor carefully tuned graph neural networks overall. GraphPFN reaches third place by Elo when more labels are available, yet its relative strengths vary substantially with dataset properties. Efficiency comparisons further qualify the benefits of pretrained reuse: GVT and GraphPFN appear on the Pareto frontiers when supervised methods are represented by their default and fully tuned configurations. Adding intermediate tuning budgets removes this advantage for GVT and leaves GraphPFN extending the estimated frontier in the label-rich setting alone. Thus, reusing pretrained parameters does not yet provide a broadly reliable route to either stronger predictions or cheaper workflows. We release the evaluation pipeline, run-level records, and an open leaderboard at https://github.com/nums-ai/nodeground.

Figures & tables

Figure 1: Top ten methods under two label splits, ranked over 51 datasets using binary AUROC and multiclass accuracy. Gray points show dataset ranks; lines show IQRs.

	Dataset coverage	Comparison control		Eval. robustness		Cost accounting	Transparency
Benchmark	Number of datasets	HPO control	Val. selection	Data splits	Multi- metric	Compute cost	Raw results	Leader- board
Yang et al. (2016) †	4	$\bm{\times}$	❍	1	$\bm{\times}$	$\bm{\times}$ $\bm{\times}$ $\bm{\times}$	$\bm{\times}$	$\bm{\times}$
Shchur et al. (2018) †	8	❍	$\triangle$	100	$\bm{\times}$	$\bm{\times}$ $\bm{\times}$ $\bm{\times}$	$\bm{\times}$	$\bm{\times}$
Pei et al. (2020) †	9	❍	❍	10	$\bm{\times}$	❍ $\bm{\times}$ $\bm{\times}$	$\bm{\times}$	$\bm{\times}$
Platonov et al. (2026) †	4	❍	❍	1	$\bm{\times}$	❍ ❍ ❍	$\triangle$	$\bm{\times}$
You et al. (2020)	18	❍	$\triangle$	3	$\bm{\times}$	❍ $\triangle$ $\triangle$	$\bm{\times}$	$\bm{\times}$

Table 1: Comparison of node-classification benchmarks and evaluation protocols. Columns correspond to the five criteria given in Section 2 . Under Cost accounting , the three symbols indicate reporting of training, tuning, and inference costs, respectively. ❍ ,

\triangle

, and

\bm{\times}

denote full, partial, and no documented support. A dagger marks an evaluation protocol rather than a benchmark. Appendix B.6 defines the criteria and documents the evidence for each assessment.

Figure 2: Application domains of the 51 retained datasets.

Figure 3: Individual property coverage of NodeGround . Panels summarize graph structure, node features, and targets. Each point represents one dataset, with colors indicating application domains. Gray boxes show interquartile ranges and black lines show medians.

Figure 4: Joint property coverage of NodeGround . (a) Node count versus average degree. (b) Median nonzero degree

q_{50}

versus hub concentration

q_{99}/q_{50}

. (c) Feature dimension versus nonzero fraction. (d) Class-balanced feature-similarity AUROC versus adjusted homophily. Dashed lines in panel (d) mark chance-level feature separation and zero adjusted homophily.

Group

Methods

Supervised

MLP (feature-only) , GCN ( Kipf and Welling, 2016 ) , GraphSAGE ( Hamilton et al., 2017 ) , GAT ( Veličković et al., 2017 ) , APPNP ( Gasteiger et al., 2018 ) , SGC ( Wu et al., 2019 ) , GCNII ( Chen et al., 2020 ) , FAGCN ( Bo et al., 2021 ) , GPRGNN ( Chien et al., 2020 ) , LINKX ( Lim et al., 2021 ) , GATv2 ( Brody et al., 2021 ) , NodeFormer ( Wu et al., 2022 ) , SGFormer ( Wu et al., 2023 ) , NAGphormer ( Chen et al., 2022 ) , Polynormer ( Deng et al., 2024 )

Foundation

G2T-FM ( Eremeev et al., 2025a ) , GraphPFN ( Eremeev et al., 2025b ) , NodePFN ( Choi et al., 2026 ) GraphAny ( Zhao et al., 2025 ) , GVT ( Lee et al., 2025 ) , Node4All ( Lee and Yoo, 2026 )

Table 2: Methods evaluated in NodeGround . We evaluate 15 supervised models trained from scratch for each dataset and split, alongside six recent GFM .

Panel of Figure 5 — Figure 5: Overall comparison of GFMs and supervised methods. Primary-metric Elo scores (binary AUROC and multiclass accuracy) across all 51 datasets under (a) 10/10/80 and (b) 50/25/25. Orange denotes GFMs, while light and dark blue denote the default and tuned supervised methods.

Figure 6: Primary-metric ranks of all methods. From left to right: binary, 10/10/80; binary, 50/25/25; multiclass, 10/10/80; and multiclass, 50/25/25. Gray points show dataset ranks, gray horizontal lines show interquartile ranges, and large markers show mean ranks (blue circles for supervised methods and orange diamonds for GFMs).

Panel of Figure 7 — Figure 7: Primary-metric ranks across dataset conditions. GCNII, GPRGNN, and GraphPFN are compared across 22 subgroups. The radial scale is truncated at rank 15. Farther outward indicates a better (lower) rank; $N$ denotes the number of datasets in each subgroup.

Panel of Figure 8 — Figure 8: Performance–cost frontiers under both label splits. Mean Improvability is plotted against mean workflow-time overhead for GFMs and the default and tuned supervised methods. Dashed lines mark the Pareto frontiers; lower is better on both axes.

Panel of Figure 9 — Figure 9: Performance–cost trajectories across tuning budgets. Lines trace supervised methods from 1 to 201 configurations, while diamonds mark GFMs, under (a) 10/10/80 and (b) 50/25/25.

Appendix figures & tables31 assets

Supplementary material from the paper’s appendix.

Appendix

Cohort	Definition	Split	Datasets	Methods	Idx.	Cells	Imputed
V1	Full benchmark	10/10/80	51	21	5	5355	144
V1	Full benchmark	50/25/25	51	21	5	5355	244
V2	Complete case	10/10/80	28	21	5	2940	0
V2	Complete case	50/25/25	28	21	5	2940	0
V3	Restricted method	10/10/80	51	15	5	3825	10
V3	Restricted method	50/25/25	51	15	5	3825	10

Appendix

Table 3: Predictive-result cohorts used throughout the paper and release. A run cell is one dataset–method–split–index record. Cohort V3 is near-complete rather than a strict complete-case cohort. Its 20 imputed run cells are GraphAny and GVT on ogbn_arxiv .

Panel of Figure 10 — Figure 10: Elo leaderboards across analysis cohorts and label splits. Cohort V1 evaluates all 21 methods on all 51 datasets using the benchmark replacement policy; Cohort V2 uses the 28 complete-case datasets; and Cohort V3 evaluates 15 near-complete methods on all 51 datasets. Dashed outlines identify partially imputed results, and error bars show 95% confidence intervals.

Panel of Figure 11 — Figure 11: Binary-metric ranks across analysis cohorts under the 10/10/80 split. Gray points show dataset-level ranks, and gray horizontal lines show interquartile ranges. Large blue circles and orange diamonds show mean ranks for supervised methods and GFMs, respectively. Lower ranks are better.

Panel of Figure 12 — Figure 12: Binary-metric ranks across analysis cohorts under the 50/25/25 split. Panels and visual encodings follow Figure 11 .

Panel of Figure 13 — Figure 13: Multiclass-metric ranks across analysis cohorts under the 10/10/80 split. Gray points show dataset-level ranks, and gray horizontal lines show interquartile ranges. Large blue circles and orange diamonds show mean ranks for supervised methods and GFMs, respectively. Lower ranks are better.

Panel of Figure 14 — Figure 14: Multiclass-metric ranks across analysis cohorts under the 50/25/25 split. Panels and visual encodings follow Figure 13 .

Panel of Figure 15 — Figure 15: Workflow-time performance–cost frontiers under alternative analysis cohorts. Mean Improvability is plotted against Mean Overhead from total target-dataset time; lower is better on both axes. Dashed lines mark Pareto frontiers. The Cohort V1 analyses appear in Figure 8 .

Comparison	Coefficient	10/10/80	50/25/25	Interpretation
Multiclass accuracy vs. Macro-F1	Spearman	0.978	0.989	Strong agreement
Binary AUROC vs. positive precision	Kendall	0.519	0.664	Metric-sensitive ordering

Appendix

Table 4: Selected ranking correlations on Cohort V2. The released correlation table contains every metric, split, and aggregation comparison.

Split	Rank–Elo	Rank–MI	Elo–MI	Leader	Max. shift	Same top four
10/10/80	0.990	0.962	0.969	GCNII	4	No
50/25/25	0.993	0.972	0.955	GraphPFN	6	No

Appendix

Table 5: Agreement among cross-dataset aggregation rules on Cohort V2. Rank denotes mean within-dataset rank and MI denotes Mean Improvability. Each coefficient is the Spearman correlation between the indicated pair of complete method rankings. Leader is the method ranked first by all three rules. Max. shift is the largest change in any method’s position across the three rankings, and Same top four indicates whether they retain the same set of four leading methods.

Panel of Figure 16 — Figure 16: Workflow-peak-memory performance–cost frontiers on Cohort V1. Mean Improvability is plotted against Mean Overhead from the larger of fit/search and prediction peak GPU memory. Dashed lines mark Pareto frontiers; lower is better on both axes.

Panel of Figure 17 — Figure 17: Pairwise primary-metric win rates on Cohort V1. Each cell gives the percentage of the 51 datasets on which the row method beats the column method. Methods are ordered separately for each split by increasing mean within-dataset primary-metric rank.

	10/10/80		50/25/25
Method	Significant wins	Significant losses	Significant wins	Significant losses
GCNII	11	0	8	0
GraphPFN	1	0	7	0
GPRGNN	6	0	6	0
FAGCN	6	0	8	0
GCN	5	0	7	0

Appendix

Table 6: Holm-corrected pairwise comparisons. Counts are computed on the 28-dataset predictive complete-case cohort, with Holm correction applied jointly across all 210 method pairs within each split. Counts summarize the 20 paired method comparisons for each row method. A win requires a positive median paired effect and adjusted

p<0.05

Method	$\Delta$ (pp)	$B^{1\mathrm{pp}}$	Cost at $B$ ( $\times$ )	$\Delta$ (pp)	$B^{1\mathrm{pp}}$	Cost at $B$ ( $\times$ )
	10/10/80			50/25/25
APPNP	4.5	25	14.0	6.0	25	14.8
FAGCN	3.0	10	5.6	2.9	25	14.8
GAT	7.3	25	13.9	10.0	50	29.4
GATv2	6.9	25	10.2	10.0	50	20.1
GCN	5.5	50	27.7	8.2	25	14.7

Appendix

Table 7: Tuning dependence by method.

\Delta

is the test-performance gain from 1 to 201 configurations.

B^{1\mathrm{pp}}

is the smallest sampled budget within one percentage point of the 201-configuration result. Results average the 28-dataset predictive complete-case cohort, five experimental indices, and 20 nested configuration samples. Cost is the workflow multiplier relative to the default, estimated by linear interpolation between the measured default and 201-configuration endpoints.

Figure 18: How tuning improves test performance and changes the gap between validation and test gains. Left: test-score improvement over the default configuration. Right: validation improvement minus test improvement, each measured relative to the default. Positive values mean that tuning appears more beneficial on validation nodes than on test nodes. All gains are in percentage points (pp). Configurations are selected by validation loss; gains use multiclass accuracy or binary AUROC. Curves average over 28 datasets, five shared splits, and 20 search orders.

Source channel	Candidate records	Eligible records	Retained tasks
PyG	94	48	32
HF TAG	11	11	11
Heterophilous	11	7	7
OGB	1	1	1
Total	117	67	51

Appendix

Table 8: Candidate records and retained benchmark tasks by discovery channel. Eligible records pass the scope, target, feature, and scale checks. Final tasks are counted after canonicalization.

Check	Criterion	Excluded	Retained	Typical exclusions or merges
Scope	Homogeneous, single-domain, single-graph node classification.	16	101	PPI, WikipediaNetwork:Crocodile, Twitch:DE
Target	Verified categorical single label per evaluated node.	7	94	Yelp, AttributedGraphDataset:PPI, AttributedGraphDataset:Facebook
Feature	Eligible observable node attributes; no missing, identity-only, topology-only, or disallowed legacy representations.	15	79	KarateClub, Roman-empire, Minesweeper
Scale	$F\leq 10{,}000$ ; $\|V\|+\|E\|\leq 1.7\times 10^{6}$ ; $\|V\|\|E\|\leq 5\times 10^{11}$ .	12	67	NELL, Reddit, Reddit2
Canonicalization	Merge eligible loader records representing the same benchmark task.	16	51	CitationFull:Cora, CitationFull:PubMed, AttributedGraphDataset:Cora

Appendix

Table 9: Ordered candidate-screening criteria and resulting funnel. Each record is assigned to the first check that it fails.

Canonical task	Records	Merged loader records
actor	2	Actor; heterophilous_graph:actor
amazon_ratings	2	HeterophilousGraphDataset:Amazon-ratings; heterophilous_graph:amazon_ratings
chameleon	2	WikipediaNetwork:Chameleon; heterophilous_graph:chameleon
citeseer	2	Planetoid:CiteSeer; AttributedGraphDataset:CiteSeer
cora	2	Planetoid:Cora; AttributedGraphDataset:Cora
cornell	2	WebKB:Cornell; heterophilous_graph:cornell

Appendix

Table 10: Eligible loader-record aliases consolidated into canonical benchmark tasks.

Dataset	Channel	Domain	Task	Citation	License
actor	Heteroph.	Web	Actor category classification	( Pei et al., 2020 )	MIT
amazon_computer	PyG	Commerce	Product category classification	( McAuley et al., 2015 ; Shchur et al., 2018 )	MIT
amazon_photo	PyG	Commerce	Product category classification	( McAuley et al., 2015 ; Shchur et al., 2018 )	MIT
amazon_ratings	Heteroph.	Commerce	Binned mean product-rating classification	( Platonov et al., 2023b )	MIT
amherst41	PyG	Social	Student gender classification	( Traud et al., 2012 ; Lim et al., 2021 )	not stated
artnet-exp	PyG	Social	Explicit-content creator classification	( Bazhenov et al., 2026 )	Apache-2.0

Appendix

Table 11: Catalog of the 51 retained node-classification tasks. Channel labels link to upstream sources. Citations identify the corresponding data or benchmark publications. License entries reproduce terms stated by the linked upstream repository or data release. “Not stated” means that no explicit license was found at that source; it does not imply that research use is prohibited. NodeGround does not redistribute or relicense these datasets; our pipeline downloads data from upstream sources, preprocesses them locally, and reports evaluation results. The four road-network datasets derive from OpenStreetMap contributors (ODbL 1.0) ; their processed release states no separate license.

Property and definition	Subgroup Boundary #data
Number of nodes $N=\|V\|$ Number of graph nodes.	Tiny $N\leq 2{,}000$ 9 Small $2{,}000<N\leq 10{,}000$ 14 Medium $10{,}000<N\leq 100{,}000$ 19 Large $N>100{,}000$ 9
Average degree $\bar{d}=2E/N$ Mean number of neighbors per node.	Low $\bar{d}\leq 5$ 20 Medium $5<\bar{d}\leq 20$ 18 High $\bar{d}>20$ 13
Label homophily $h_{\mathrm{adj}}=(h-h_{0})/(1-h_{0})$ Label agreement along edges, adjusted for chance.	$<0$ $h_{\mathrm{adj}}<0$ 8 $0$ – $0.5$ $0\leq h_{\mathrm{adj}}<0.5$ 15 $\geq 0.5$ $h_{\mathrm{adj}}\geq 0.5$ 28
Number of features $F$ Number of columns in the node-feature matrix.	Low $F\leq 500$ 15 Medium $500<F<5{,}000$ 31 High $F\geq 5{,}000$ 5
Feature nonzero fraction $s$ Fraction of sampled finite feature entries that are nonzero.	Sparse $s\leq 0.01$ 16 Medium $0.01<s\leq 0.5$ 17 Dense $s>0.5$ 18
Number of classes $C$ Number of nonempty classes among labeled nodes.	Binary $C=2$ 11 3–10 $3\leq C\leq 10$ 32 $>10$ $C>10$ 8

Appendix

Table 12: Property definitions and subgroup boundaries for RQ2. Seven properties define the 22 subgroups in Figure 7 . Each dataset belongs to one subgroup per property; #data gives the subgroup size.

Quantity	Definition
Edge count	$E$ : number of unique undirected non-self-loop edges. Reciprocal orientations are counted once.
Density	$\rho=2E/[N(N-1)]$ for $N\geq 2$ ; zero for $N<2$ .
Degree quantiles	$q_{50}$ and $q_{99}$ : median and 99th percentile of nonzero node degrees.
Hub concentration	$q_{99}/q_{50}$ : ratio of the 99th-percentile to the median nonzero degree.
Feature-similarity AUROC	AUROC of cosine feature similarity for distinguishing same-label from different-label node pairs. Sampling is specified below.
Labeled-node count	$N_{L}$ : number of nodes with valid target labels.

Appendix

Table 13: Additional statistics for describing dataset coverage. These describe dataset coverage but do not define the RQ2 subgroups.

Dataset	$N$	Group	$\bar{d}$	Group	$h_{\mathrm{adj}}$	Group
	Number of nodes		Average degree		Label homophily
actor	7,600	Small	7.02	Medium	0.003	$0$ – $0.5$
amazon_computer	13,752	Medium	35.76	High	0.682	$\geq 0.5$
amazon_photo	7,650	Small	31.13	High	0.785	$\geq 0.5$
amazon_ratings	24,492	Medium	7.60	Medium	0.140	$0$ – $0.5$
amherst41	2,235	Small	81.39	High	0.060	$0$ – $0.5$

Appendix

Table 14: Structural values and subgroup assignments. Each property has a value column and a group column. Groups follow Table 12 ; homophily group

0

–

0.5

means

0\leq h_{\mathrm{adj}}<0.5

. The displayed values are rounded.

Dataset	$F$	Group	$s$	Group	$C$	Group	$r$	Group
	Number of features		Nonzero fraction		Classes		Class imbalance
actor	932	Medium	0.006	Sparse	5	3–10	2.30	Medium
amazon_computer	767	Medium	0.343	Medium	10	3–10	17.73	High
amazon_photo	745	Medium	0.344	Medium	8	3–10	5.86	Medium
amazon_ratings	300	Low	1.000	Dense	5	3–10	8.49	Medium
amherst41	1,193	Medium	0.005	Sparse	2	Binary	1.00	Low

Appendix

Table 15: Feature and target values and subgroup assignments. Each property has a value column and a group column.

F

is the number of features,

s

is the feature nonzero fraction,

C

is the number of classes, and

r

is the class-imbalance ratio. Groups follow Table 12 ; displayed values are rounded.

Dataset	$E$	$q_{50}$	$q_{99}$	Hub	Density	Feature AUROC	$N_{L}$	Label entropy
actor	26,659	4.00	51.00	12.75	9.23e-04	0.503	7,600	0.977
amazon_computer	245,861	23.00	245.00	10.65	2.60e-03	0.542	13,752	0.813
amazon_photo	119,081	22.00	201.32	9.15	4.07e-03	0.566	7,650	0.926
amazon_ratings	93,050	5.00	31.00	6.20	3.10e-04	0.502	24,492	0.877
amherst41	90,954	70.00	297.98	4.26	3.64e-02	0.500	2,032	1.000
artnet-exp	280,348	2.00	135.00	67.50	2.21e-04	0.521	50,405	0.469

Appendix

Table 16: Additional dataset statistics. These quantities describe dataset coverage but do not define the RQ2 subgroups.

E

counts undirected non-self-loop edges,

q_{50}

and

q_{99}

are nonzero-degree quantiles, Hub is

q_{99}/q_{50}

, and

N_{L}

counts labeled nodes. Feature AUROC and label entropy follow Table 13 .

Method	Family	Implementation used	Configuration set
MLP	Feature-only	Benchmark implementation (PyTorch)	default + 200 sampled
APPNP	GNN	Benchmark wrapper using PyG APPNP	default + 200 sampled
FAGCN	GNN	Benchmark wrapper using PyG FAConv	default + 200 sampled
GAT	GNN	Benchmark wrapper using PyG GATConv	default + 200 sampled
GATv2	GNN	Benchmark wrapper using PyG GATv2Conv	default + 200 sampled
GCN	GNN	Benchmark wrapper using PyG GCNConv	default + 200 sampled

Appendix

Table 17: Implementation provenance and evaluated configuration sets. Benchmark implementation and benchmark reimplementation denote code maintained within NodeGround rather than execution of the original authors’ repositories. Benchmark wrapper denotes NodeGround code built from the named PyG operator, whereas the GFM rows explicitly distinguish official implementations from adapters that load official checkpoints.

Method	Artifact and source	Frozen / trainable	Target operation
G2T-FM	Official G2T-FM implementation with the TabPFNv2 classifier checkpoint	TabPFNv2 frozen; no trainable head	Constructs graph-derived tabular features and supplies training labels as context. Uses n_epochs=0 and no fine-tuning.
GraphAny	Official graph_any_arxiv.pt checkpoint	Fusion MLP frozen; no trainable head	Forms feature and propagated channels, analytically solves label-conditioned linear channels from the training labels, and fuses their logits.
GraphPFN	Official GraphPFN 1.3 release	Complete pretrained model frozen; no trainable head	Runs graph-aware label-conditioned inference and selects between PCA disabled and 64 dimensions.
GVT	Official RGVT seed_42.pth checkpoint	RGVT encoder frozen; target MLP head trained	Computes representations at depths 1–8 and trains a separate validation-selected head at each depth.
Node4All	Official cgt_enc.pth checkpoint	CGT encoder frozen; target MLP head trained	Computes one canonical representation and trains a validation-selected target-specific MLP head.
NodePFN	Official v1.0.0 checkpoint_epoch_30.ckpt	Complete pretrained model frozen; no trainable head	Uses training labels as context and selects dimensionality reduction, component count, smoothing, and ensemble settings.

Appendix

Table 18: Pretrained artifacts and target-dataset adaptation. The table identifies the released artifact actually evaluated and states which components, if any, are trained on each target dataset.

Method(s)	Default	Varied values
Shared supervised training	Adam; learning rate $10^{-2}$ , weight decay $5\!\times\!10^{-4}$ , 2,500 epochs, patience 100	learning rate $\{10^{-3},5\!\times\!10^{-3},10^{-2}\}$ ; weight decay $\{0,5\!\times\!10^{-5},5\!\times\!10^{-4}\}$ ; patience $\{20,100\}$
APPNP	hidden 64; feature dropout .5; $K=10$ ; $\alpha=.1$ ; propagation dropout 0	hidden $\{64,128,256,512\}$ ; feature dropout $\{0,.3,.5,.7\}$ ; $K\in\{5,10,20\}$ ; $\alpha\in\{.1,.2,.5\}$ ; propagation dropout $\{0,.3,.5\}$ ; shared training
FAGCN	hidden 32; feature dropout .5; 2 layers; $\epsilon=.3$ ; edge dropout .5	hidden $\{32,64,128\}$ ; feature and edge dropout $\{0,.3,.5,.7\}$ ; layers $\{1,2,3,4\}$ ; $\epsilon\in\{.1,.2,.3,.5\}$ ; shared training
GAT, GATv2	hidden 64; one head; feature dropout .5; attention dropout .2; 2 layers; no normalization, residual, or pre-transform	hidden $\{64,128,256,512\}$ ; heads $\{1,4\}$ ; feature dropout $\{0,.3,.5,.7\}$ ; attention dropout $\{0,.2\}$ ; layers $\{1,2,4,8\}$ ; normalization $\{\texttt{none},\texttt{layernorm}\}$ ; residual and pre-transform booleans; shared training
GCN, GraphSAGE	hidden 64; feature dropout .5; 2 layers; no normalization, residual, or pre-transform	hidden $\{64,128,256,512\}$ ; feature dropout $\{0,.3,.5,.7\}$ ; layers $\{1,2,4,8\}$ ; normalization $\{\texttt{none},\texttt{layernorm}\}$ ; residual and pre-transform booleans; shared training
GCNII	hidden 64; feature dropout .6; 16 layers; $\alpha=.1$ ; $\theta=.5$ ; shared weights; no normalization, residual, or pre-transform	hidden $\{64,128,256,512\}$ ; feature dropout $\{0,.3,.5,.7\}$ ; layers $\{4,8,16,32\}$ ; $\alpha\in\{.1,.2,.5\}$ ; $\theta\in\{.5,1,1.5\}$ ; normalization, residual, and pre-transform choices; shared training

Appendix

Table 19: Default configurations and source-defined search spaces. The shared training row applies where a method row does not give a different training setting. NodeFormer and Polynormer learning rates, weight decays, and dropout values are categorical choices, not continuous or log-uniform ranges.

Method	Trial records	Cells	Default selected	Median ID	Modal selected settings
APPNP	95,403/102,510	510/510	2.5%	101.5	hidden 512; feature drop .7; $K=5$ ; $\alpha=.5$ ; propagation drop 0
FAGCN	102,510/102,510	510/510	0.2%	104.5	hidden 128; feature drop .7; 4 layers; $\epsilon=.5$ ; edge drop .3
GAT	102,510/102,510	510/510	0.2%	97	hidden 64; 4 heads; feature/attention drop .3/.2; 1 layer; layer norm; residual on; pre-transform off
GATv2	92,862/92,862	462/510	0.0%	99	hidden 64; 4 heads; feature/attention drop .3/.2; 1 layer; layer norm; residual on; pre-transform off
GCN	102,510/102,510	510/510	1.0%	102	hidden 512; feature drop .3; 2 layers; layer norm; residual on; pre-transform off
GCNII	100,500/100,500	500/510	3.4%	99.5	hidden 512; feature drop .7; 32 layers; $\alpha=.5$ ; $\theta=.5$ ; layer norm; residual and pre-transform off

Appendix

Table 20: Optimization execution and selected-configuration summaries. Trial records are successful/attempted totals. Cell coverage is measured against 510 possible dataset–split–index cells per method. Default selected is the percentage of successful cells whose selected trial identifier is 1. For one-configuration methods this is necessarily 100%. The last column gives the modal selected values of method-specific search dimensions.

task	primary	secondary	selection loss
Multiclass	Accuracy	Macro-F1; cross-entropy	Validation cross-entropy
Binary	AUROC	Accuracy; class-1 precision; binary cross-entropy	Validation binary cross-entropy

Appendix

Table 21: Primary and secondary predictive metrics. Selection always minimizes validation cross-entropy or binary cross-entropy; primary denotes the headline test metric, not the selection objective.

Dataset	Task	Best, 10/10/80	Best, 50/25/25
actor	M (Acc.)	GraphPFN 38.15 $\pm$ 0.56	GraphPFN 39.64 $\pm$ 0.34
amazon_computer	M (Acc.)	GraphPFN 91.09 $\pm$ 0.27	GraphPFN 93.04 $\pm$ 0.36
amazon_photo	M (Acc.)	GraphPFN 94.95 $\pm$ 0.33	GraphPFN 96.25 $\pm$ 0.47
amazon_ratings	M (Acc.)	GraphPFN 43.78 $\pm$ 0.29	LINKX 50.63 $\pm$ 1.07
amherst41	B (AUROC)	LINKX 76.50 $\pm$ 1.36	LINKX 89.42 $\pm$ 1.73
artnet-exp	B (AUROC)	GraphPFN 84.60 $\pm$ 0.33	GraphPFN 86.75 $\pm$ 0.27

Appendix

Table 22: Per-dataset primary-metric leaders in the 51-dataset cohort. Leaders are selected among the single report point of each method: the validation-selected tuned configuration for supervised methods and the canonical configuration for GFMs. Values are mean

\pm

standard deviation over five shared partitions. M uses multiclass accuracy and B uses binary AUROC. No imputed result is selected as a leader.

Configuration	Mean rank 10/10/80	Mean rank 50/25/25	Rank 10/10/80	Rank 50/25/25
GCNII [hp_tuned]	6.4216	7.1667	1	1
GPRGNN [hp_tuned]	6.8039	7.3333	2	2
FAGCN [hp_tuned]	8.2353	8.4608	3	3
GCN [hp_tuned]	9.9216	9.1373	4	4
GAT [hp_tuned]	10.8431	10.7157	5	9
GraphSAGE [hp_tuned]	11.2843	9.8627	6	7

Appendix

Table 23: Primary-metric ranks of all 36 default, tuned, and canonical configurations in Cohort V1.

Method	Observed 10/10/80	Imputed 10/10/80	Observed 50/25/25	Imputed 50/25/25
GATv2	46	5	46	5
GCNII	50	1	50	1
NAGphormer	50	1	50	1
NodeFormer	41	10	41	10
G2T-FM	44	7	30	21
GraphAny	50	1	50	1

Appendix

Table 24: Methods requiring at least one GCN-Default replacement in Cohort V1. All unlisted methods have native results on all 51 datasets under both label ratios.

Explore similar work

Jun 23, 2026cs.LG

A Fair Evaluation of Graph Foundation Models for Node Property Prediction

Due to the wide use of graph-structured data in different fields of industry and science, the development of Graph Foundation Models (GFMs) has recently attracted a lot of attention. While many different types of models are called GFMs, particular interest has been paid to GFMs designed for node property prediction tasks, which is one of the most popular settings in Graph ML with lots of real-world applications from fraud detection in financial and social networks to recommendation systems for e-commerce and user-generated content platforms. While a number of GFMs for this task have been recently proposed, the field has not converged to a unified evaluation setting, and different works evaluate their models in widely different ways, preventing reliable comparison of GFMs with each other and with other types of models. In this work, we conduct a fair and rigorous reevaluation of 9 recent GFMs for node property prediction, comparing them to strong Graph Neural Network (GNN) baselines. We find that, among these GFMs, only the most recent ones based on the Prior-data Fitted Networks paradigm outperform well-tuned GNNs in predictive performance, although at a higher inference cost.

Oleg Platonov, Gleb Bazhenov, Dmitry Eremeev +1

1HSE University · 2Yandex Research

Graph Foundation Model Graph Neural Networks Tabular Prior-Data Fitted Network Foundation Model Algorithmic Fairness

Ask PDF

Sep 25, 2025cs.LG

GraphPFN: A Prior-Data Fitted Graph Foundation Model

Graph foundation models face several fundamental challenges including transferability across diverse domains and data scarcity, which calls into question the very feasibility of creating such models. However, despite similar challenges, the tabular domain has recently witnessed the emergence of the first successful foundation models such as TabPFN. These models are based on the prior-data fitted networks (PFN) framework, in which models are pretrained on carefully designed synthetic datasets to make predictions in an in-context learning setting. Recently, G2T-FM, a framework that converts graph node-level tasks into tabular tasks, has made the first step towards adopting PFNs for graphs, yet it is limited to hand-crafted features and was never pretrained on graph data. In this work, we make the next step by proposing GraphPFN, a PFN-based model designed and pretrained specifically for graph node-level tasks. Following the PFN framework, we first design a prior distribution of synthetic attributed graphs by using a novel combination of multi-level stochastic block models and a preferential attachment process for structure generation and graph-aware structured causal models for attribute generation. Then, we augment the tabular foundation model LimiX with attention-based graph neighborhood aggregation layers and train it on millions of synthetic graphs sampled from our prior. On diverse real-world graph datasets with node-level tasks, GraphPFN achieves state-of-the-art results in both in-context learning and finetuning regimes, outperforming G2T-FM, prior GFMs, and task-specific GNNs trained from scratch. More broadly, GraphPFN shows the potential of PFN-based models for building graph foundation models.

Dmitry Eremeev, Oleg Platonov, Gleb Bazhenov +2

1HSE University · 2Yandex Research

Graph Foundation Model Graph Neural Networks Tabpfn Graph Representations

Ask PDF

Jul 19, 2026cs.LG

Node4All: Learning Node Representation Beyond Datasets

Node representation learning has advanced rapidly, yet most existing methods rely on per-dataset training and hyperparameter tuning. This dataset-specific optimization comes from the difficulty of designing reusable graph models that generalize across diverse graph datasets. In this work, we introduce Node4All, a node representation learner applicable to arbitrary graph datasets without any dataset-specific optimization. Node4All is built on two complementary ideas. At the architectural level, we introduce the Channel Graph Transformer (CGT), which enables a single fixed parameterization to process arbitrary graph datasets. At the learning level, we propose a self-supervised learning based on a series of synthetic graphs. Together, these components enable generalization beyond individual datasets, which is infeasible with existing architectures and learning frameworks. We extensively evaluate Node4All on node classification across 25 benchmarks against 21 baselines, covering both supervised and self-supervised methods. Despite all baselines being trained and optimized for each dataset, a single Node4All, applied uniformly across the datasets, achieves a competitive ranking of 5th among 21 baselines. Moreover, Node4All supports one-shot and in-context learning with an appropriate predictor and outperforms recent graph foundation models (GFMs) in these settings. These results demonstrate that Node4All not only achieves reusability across arbitrary graph datasets, but also remains an effective solution in practice. Code and model checkpoints are available in https://github.com/dooho00/node4all.

Dooho Lee, Jaemin Yoo

KAIST · Daejeon, Republic of Korea · Seoul National University +1

Graph Representation Learning Graph Representations

Ask PDF