The Universal Classifier for Graph Learning

Ben Finkelshtein; André Linhares; Petar Veličković; Bryan Perozzi; Mikhail Galkin

cs.LGSep 28, 2026

The Universal Classifier for Graph Learning

Authors: Ben Finkelshtein, André Linhares, Petar Veličković, Bryan Perozzi, Mikhail Galkin

Organizations: Google Research · Google DeepMind

Abstract

While foundation models have revolutionized natural language processing and computer vision by leveraging universal vocabularies, Graph Machine Learning (GML) remains fractured due to the absence of a unified feature and structural representation across diverse domains. Existing works claiming to be Graph Foundation Models (GFMs) are typically restricted to node-level predictions or require fixed feature dimensions, failing to provide a truly task-agnostic backbone for the full spectrum of graph learning applications. In this paper, we introduce the Universal Classifier (UC), which supports arbitrary feature and class cardinalities, unifying node-, edge-, and graph-level objectives under a single similarity-based classification objective. The UC reformulates all node-, edge-, and graph-level prediction tasks as maximizing similarity in the latent space: by lifting heterogeneous features and labels into 3D latent tensors, the model learns transferable features independent of specific input schemas. This architecture allows a single pre-trained model to generalize to node classification, node regression, and link prediction across unseen graphs with varying feature semantics. Experiments show strong zero-shot transfer performance across node-, link-, and graph-level tasks.

Figures & tables

Panel of Figure 1 — Figure 1: Overview of the Universal Classifier generalizing across tasks. (a) Node Classification and Regression: a query node is compared to a support set of anchors, with similarities aggregated by class. In regression, continuous targets are discretized into bins, treated as distinct classes, and recovered via similarity-weighted interpolation. (b) Link Prediction: a source node acts as a query and is evaluated against target nodes. Nodes connected to the query serve as anchors of a positive class, while all other nodes act as anchors of a negative class. (c) Graph Classification and Regression: a source graph is represented by pooling its respective features into a single global query, which is then compared against anchor graphs representing specific labels, similar to node-level tasks.

Panel of Figure 3 — Figure 3: The UC performance across tasks: (a) binary (left, AP $\uparrow$ ) and multi-class (right, Accuracy $\uparrow$ ) node classification; (b) node (left, $R^{2}$ $\uparrow$ ) and graph (right, RSE $\downarrow$ ) regression; and (c) link prediction (Hits@10 $\uparrow$ ), comparing zero-shot transfer and fine-tuning to tabular FMs and supervised baselines.

Parameter	roman- empire (Acc)	city-roads- M-RH ( $R^{2}$ )
Base UC	78.61	74.08
- 2 Trf layers	75.63	70.90
- 3 hops	77.25	67.94
- GNN encoder	64.35	53.30
- Tabular encoder	74.27	73.11
- Mean score agg	72.83	71.79

Table 1: UC architectural ablations.

Panel of Figure 4 — Figure 4: t-SNE projections of UC anchor representations for node classification ( roman-empire, arxiv ) and regression ( artnet-views-RH ) datasets cluster around the labels. In regression datasets, labels are ordinal bins constructed around centroids, and nearby bins cluster together.

Figure 5: Inference-time performance (accuracy for classification and

R^{2}

for regression) scaling with larger context sizes. 5% of anchors is often enough to recover 80+% of the final performance.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Task	Dataset	Graphs	Nodes	Edges	Features	Classes
Node Classification	amazon-ratings	1	24,492	93,050	300	5
	artnet-exp	1	50,405	280,348	75	2
	arxiv	1	169,343	1,157,799	128	40
	city-reviews	1	148,801	1,165,415	37	2
	full-cora	1	19,793	63,421	8,710	70
	full-dblp	1	17,716	52,867	1,639	4

Appendix

Table 2: Summary statistics of the datasets used in our experiments. Node and edge counts are averaged per graph, and edges are counted after symmetrization, matching the graph representation used at runtime. The link-prediction datasets are reported separately in Table 3 . EC graphs carry no node features, so we turn node types into one-hot encoded features. For the EC datasets, the number of graphs is given per train/validation/test split, and the EFF and VOUT targets share the same graphs.

Dataset	Nodes	Undirected edges	Message-passing edges	Features	Mean degree
roman-empire	22,662	32,927	30,927	300	2.9
wiki-cs	11,701	215,603	213,603	300	36.9
city-roads-M	57,073	107,104	105,104	68	3.8
tolokers-2	11,758	519,000	517,000	19	88.3

Appendix

Table 3: Statistics of the link prediction datasets after the preprocessing described above. Message-passing edges excludes the

2{,}000

held-out validation and test edges. Feature counts are post-transform (one-hot encoding and imputation), so city-roads-M and tolokers-2 exceed the raw GraphLand counts.

Model	amazon-ratings	full-cora	full-dblp	roman-empire	arxiv
Supervised GNN	47.18 $\pm$ 0.42	58.95 $\pm$ 0.55	73.87 $\pm$ 1.35	43.93 $\pm$ 0.45	73.65 $\pm$ 0.11
Zero-Shot Transfer
GraphAny	42.80 $\pm$ 0.09	51.18 $\pm$ 0.78	71.48 $\pm$ 1.44	63.34 $\pm$ 0.58	58.85 $\pm$ 0.03
TS-Mean	42.27 $\pm$ 1.40	53.58 $\pm$ 0.73	66.42 $\pm$ 3.65	66.36 $\pm$ 1.02	56.33 $\pm$ 2.58
NodePFN	44.68 $\pm$ 0.48	-	74.71 $\pm$ 0.39	-	-
UC	52.83	68.64	85.10	78.61	65.43

Appendix

Table 4: Node classification accuracies.

Model	hm-categories (Acc)		tolokers-2 (AP)		city-reviews (AP)		artnet-exp (AP)
	RL	RH	RL	RH	RL	RH	RL	RH
Supervised GNN	67.96 $\pm$ 0.33	79.19 $\pm$ 0.21	53.78 $\pm$ 1.34	63.76 $\pm$ 1.30	77.67 $\pm$ 0.13	81.10 $\pm$ 0.11	46.62 $\pm$ 0.32	50.62 $\pm$ 0.35
Zero-Shot Transfer
TS-Mean	20.09 $\pm$ 1.29	15.48 $\pm$ 1.93	38.54 $\pm$ 0.94	31.83 $\pm$ 2.55	43.46 $\pm$ 5.17	-	20.44 $\pm$ 1.05	13.38 $\pm$ 2.59
GraphPFN	-	-	60.33 $\pm$ 0.60	-	79.89 $\pm$ 0.10	-	50.85 $\pm$ 0.31	-
UC	48.69	54.83	67.18	69.26	86.53	86.57	65.46	66.92

Appendix

Table 5: Node classification accuracies on GraphLand datasets.

Model	city-roads-M		city-roads-L		twitch-views		artnet-views
Model	RL	RH	RL	RH	RL	RH	RL	RH
End-to-end
XGBoost	59.14 $\pm$ 0.11	70.93 $\pm$ 0.05	53.75 $\pm$ 0.07	64.62 $\pm$ 0.07	13.39 $\pm$ 0.00	13.34 $\pm$ 0.02	32.74 $\pm$ 0.04	36.83 $\pm$ 0.06
Supervised GNN	59.11 $\pm$ 0.20	70.53 $\pm$ 0.40	53.43 $\pm$ 0.20	66.03 $\pm$ 0.24	72.93 $\pm$ 0.17	76.06 $\pm$ 0.30	53.36 $\pm$ 0.78	59.01 $\pm$ 0.52
Zero-Shot Transfer
GraphPFN	63.55 $\pm$ 0.29	-	-	-	73.27 $\pm$ 0.13	-	61.77 $\pm$ 0.23	-

Appendix

Table 6: Node regression

R^{2}

on GraphLand datasets.

Model	roman-empire	wiki-cs	city-roads-M	tolokers-2
End-to-end
SEAL	87.33 $\pm$ 1.15	86.53 $\pm$ 1.97	99.47 $\pm$ 0.61	79.73 $\pm$ 0.61
NBFNet	88.00 $\pm$ 0.01	92.00 $\pm$ 0.40	99.20 $\pm$ 0.01	92.53 $\pm$ 0.92
BUDDY	58.40 $\pm$ 11.09	83.60 $\pm$ 4.51	91.47 $\pm$ 5.43	72.93 $\pm$ 1.22
Zero-Shot Transfer
UC	72.00	87.60	99.60	74.00

Appendix

Table 7: Link prediction Hits@10 on the four link prediction datasets.

	EC-5		EC-7		EC-10
Method	EFF	VOUT	EFF	VOUT	EFF	VOUT
End-to-end
Ridge ( $\alpha{=}1.0$ )	0.760	0.760	0.970	0.897	0.893	0.948
KNN ( $k{=}10$ )	0.778	0.783	1.009	1.001	0.935	1.053
Random Forest (200, $d{=}5$ )	0.752	0.751	0.988	0.926	0.889	0.952
GBDT (50, $d{=}3$ )	0.720	0.724	0.989	0.890	0.877	0.977

Appendix

Table 8: Graph Regression RSE on Electronic Circuits datasets.

Graph Representation Learning

Graph representation learning develops methods for encoding nodes, edges, and whole graphs for prediction, comparison, and transfer.

See all 305 papers

To Learn is to Wander: Learning Across Graphs and Tasks with Random Walks
Oct 5, 2026Louis Tichelman, Xingyue Huang, Jinwoo Kim +1Multi-Task Learning Graph Transfer Learning
TU Wien · AITHYRA · University of Oxford +1
Mochi: Aligning Pre-training and Inference for Efficient Graph Foundation Models via Meta-Learning
Apr 23, 2026João Mattos, Arlei SilvaGraph Representation Learning Meta-Learning
Computer Science Department · Ken Kennedy Institute
Bridging Input Feature Spaces Towards Graph Foundation Models
May 6, 2026Moshe Eliasof, Krishna Sri Ipsit Mantri, Beatrice Bevilacqua +2Graph Representation Learning Transfer Learning
University of Cambridge · Ben-Gurion University of the Negev · Purdue University

Oct 5, 2026J/K move · Enter open · S save

To Learn is to Wander: Learning Across Graphs and Tasks with Random Walks

Louis Tichelman, Xingyue Huang, Jinwoo Kim, İsmail İlkan Ceylan

TU Wien · AITHYRA · University of Oxford+1

Multi-Task Learning Graph Transfer Learning Transfer Learning Graph Foundation Models KG Completion

Open paper AskPDF arXiv

Graph foundation models aim to transfer across graphs, feature spaces, relational schemas, and prediction tasks, yet existing approaches typically generalize only within particular graph modalities or tasks. We propose Wander, a graph foundation model designed to operate across these settings within a single pretrained checkpoint. Following the prior-predictive perspective, we formulate graph learning as completion of a partially observed graph. We realize this task-general view through a common interface based on random walks, allowing the same model to operate across homogeneous and multi-relational graphs with varying features, labels, and relational schemas. Wander can increase its structural context at inference time without changing its learned parameters and, under suitable assumptions, universally approximates the corresponding Bayes-optimal predictor on bounded connected graphs. Empirically, a single pretrained checkpoint achieves state-of-the-art or highly competitive results across node classification, homogeneous link prediction, and knowledge-graph link prediction. Moreover, joint pretraining across graph modalities and tasks preserves performance in specialized settings while enabling positive transfer and the composition of separately learned capabilities at inference time.

Abstract

Graph Representation Learning

Explore similar work

To Learn is to Wander: Learning Across Graphs and Tasks with Random Walks

Mochi: Aligning Pre-training and Inference for Efficient Graph Foundation Models via Meta-Learning

Bridging Input Feature Spaces Towards Graph Foundation Models