Of Graphs and Tables: Zero-Shot Node Classification with Tabular Foundation Models

Adrian Hayler, Xingyue Huang, Ismail Ilkan Ceylan, Michael M. Bronstein, Ben Finkelshtein

TL;DR

Rather than training complex Graph Foundation Models (GFMs) on scarce and noisy real-world graph data, TabGFM reformulates zero-shot node classification as a tabular learning task. By converting graph structure and features into tabular columns and running inference with pretrained Tabular Foundation Models (like TabPFN), TabGFM outperforms existing GFMs and task-specific GNNs by over 7% across 28 diverse benchmarks—without requiring any graph pretraining.

What Problem Does This Paper Address?

Graph Foundation Models aim to generalize zero-shot across unseen graphs with arbitrary features and label spaces (often evaluated under the label inpainting setting). However, progress has stalled: current GFMs rarely outperform standard, task-specific Graph Neural Networks (GNNs) trained from scratch.

The core bottleneck is data. Most available graph pretraining datasets are small, outdated, or synthetically artificial, failing to reflect real-world topology. In contrast, Tabular Foundation Models (TFMs)—such as TabPFN—are trained on hundreds of millions of synthetically generated causal models, instilling powerful general-purpose in-context learning capabilities. The authors ask: Can we bypass the graph data bottleneck entirely by reinterpreting node classification as a tabular in-context learning problem?

Methods

Once structural information is extracted via neighborhood smoothing or positional encodings, predicting node labels reduces to a classic tabular problem: mapping a matrix of row feature vectors to target labels.

TabGFM operationalizes this through a lightweight, training-free, three-stage pipeline:

  1. Node-Level Encoding (Graph $\to$ Table): Each node becomes a row. Features are generated via:
  2. Feature Encoders: Raw node features $X$, multi-hop neighborhood-smoothed averages ($A^k X$ for $k \in \{1,\dots,4\}$), and closed-form LinearGNN representations.
  3. Structural Encoders: Random Walk Positional Encodings (RandomWalkPE), Laplacian Eigenvector Positional Encodings (LaplacianEigenvectorPE), and frozen Graph Positional Structural Embeddings (GPSE).
  4. Size-Constrained Tabular Subsampling: TabPFN accepts at most 10,000 rows and 500 columns, but encoded graphs can have up to 90,000 labeled nodes and 40,000 features. TabGFM builds $B$ subsampled tables by retaining all unlabeled query rows, uniformly sampling feature/structure columns (a 300:100 ratio), and performing class-balanced subsampling of labeled context rows. Multi-class settings with $>10$ classes are handled via Error-Correcting Output Codes (ECOC).
  5. Ensemble Aggregation: Predictions across the $B$ subsampled TabPFN models and 8 LinearGNN predictors are combined using Caruana’s greedy ensemble selection algorithm, optimizing weights on a held-out set of labeled anchor nodes.
       [ Input Graph ]
              │
  ┌───────────┴───────────┐
  ▼                       ▼
Feature Encoders     Structure Encoders
(Raw, A^k X, LinGNN) (RWPE, LapPE, GPSE)
  │                       │
  └───────────┬───────────┘
              ▼
   [ Wide Node Table Z ]
              │
    (Subsample B Tables)
              │
              ▼
     [ B x TabPFN Models ]  +  [ 8 x LinearGNNs ]
              │                       │
              └───────────┬───────────┘
                          ▼
            [ Greedy Ensemble Selection ]
                          ▼
             Final Node Predictions

Datasets

The authors evaluate TabGFM on 28 diverse real-world node classification benchmarks spanning varying domains, scales, homophily levels, and label counts:

  • Citation networks: Cora, CiteSeer, PubMed, Full-Cora, Full-DBLP, OGB-Arxiv.
  • Web & Social networks: Chameleon, Squirrel, Actor, Roman-Empire, Deezer, LastFM-Asia, BlogCatalog, Wiki, Wiki-CS, Wiki-Attr.
  • E-commerce & Co-authorship: Amazon-Ratings, Computers, Photo, Co-CS, Co-Physics.
  • Heterophilous / Tabular / Synthetic graphs: Cornell, Texas, Wisconsin, Brazil, USA, Europe, Minesweeper, Questions, Tolokers.

No new datasets were introduced; official benchmark splits were used throughout.

Key Insights and Ideas

  • Modality Reframing Trumps Complex Architectures: Instead of inventing ever-more-complex message-passing schemes or pretraining architectures on scarce graph datasets, mapping graph data into a modality with mature foundation models (tables) provides stronger inductive biases.
  • Off-the-shelf TFMs Generalize to Graphs Zero-Shot: TabPFN, trained exclusively on synthetic causal tables without a single graph, effectively solves node classification out of the box when provided with neighborhood and positional feature encodings.
  • Architectural Diversity Enables Strong Ensembles: While LinearGNNs have lower standalone accuracy than TabPFN, their distinct inductive biases make their errors less correlated. Combining TabPFN with LinearGNNs via greedy ensemble selection significantly outperformed simple uniform averaging (73.26% vs. 71.61%).

Key Results

  • Outperforms State-of-the-Art GFMs: Across all 28 datasets, TabGFM achieves an average accuracy of 73.10%, outperforming existing GFMs GraphAny (65.56%) and TS-Mean (65.78%) by over 7%.
  • Beats Supervised End-to-End Baselines: TabGFM surpasses an end-to-end trained GAT (66.55%) by more than 6.5% in a strictly zero-shot transfer setting.
  • Scales with Ensemble Size: Performance steadily increases as the number of subsampled tables ($B$) grows from 1 to 10 (from 72.20% at $B=1$ to 73.26% at $B=10$), demonstrating that diverse random column/row subsamples provide complementary structural views.
  • Outperforms Multi-Graph Pretrained Baselines: When TS-Mean was pretrained on 9 graph datasets, TabGFM still outperformed it on the remaining 20 datasets (74.39% vs. 68.57%), winning on 16 of the 20 benchmarks.

Limitations and Future Work

  • Task Scope: TabGFM is currently restricted to node classification (label inpainting) and does not natively support link prediction or graph-level classification.
  • Handcrafted Encoders: The pipeline depends on fixed, predefined feature and structural encoders (e.g., GPSE, LapPE, diffusion smoothing) rather than learning graph representations end-to-end.
  • TFM Context Window Limits: Because current TFMs cannot ingest large context tables directly, the method relies on subsampling and ensembling heuristics, which can introduce redundancy on small datasets and add inference latency.
  • Future Directions: Adapting TFMs to link prediction, fine-tuning tabular models directly on graph-derived tables, and expanding context lengths to process entire graphs in a single forward pass.

Summary updated

Abstract

Graph foundation models (GFMs) have recently emerged as a promising paradigm for achieving broad generalization across various graph data. However, existing GFMs are often trained on datasets that were shown to poorly represent real-world graphs, limiting their generalization performance. In contrast, tabular foundation models (TFMs) not only excel at classical tabular prediction tasks but have also shown strong applicability in other domains such as time series forecasting, natural language processing, and computer vision. Motivated by this, we take an alternative view to the standard perspective of GFMs and reformulate node classification as a tabular problem. Each node can be represented as a row with feature, structure, and label information as columns, enabling TFMs to directly perform zero-shot node classification via in-context learning. In this work, we introduce TabGFM, a graph foundation model framework that first converts a graph into a table via feature and structural encoders, applies multiple TFMs to diversely subsampled tables, and then aggregates their outputs through ensemble selection. Through experiments on 28 real-world datasets, TabGFM achieves consistent improvements over task-specific GNNs and state-of-the-art GFMs, highlighting the potential of tabular reformulation for scalable and generalizable graph learning.
BibTeX
@inproceedings{hayler2025graphs,
  author = {Adrian Hayler and Xingyue Huang and Ismail Ilkan Ceylan and Michael M. Bronstein and Ben Finkelshtein},
  title = {{Of Graphs and Tables: Zero-Shot Node Classification with Tabular Foundation Models}},
  booktitle = {NPGML @ NeurIPS},
  year = {2025},
  url = {https://openreview.net/forum?id=u3vyZf5Jv2},
}
Figure