Learning Universal Topological Vocabularies for Graph Transfer

307 concepts235 papers3 proposals

In fields like natural language processing and computer vision, foundation models transfer effortlessly across tasks because words and pixels share a universal format. Graphs, however, are notoriously siloed. A citation network shares zero feature overlap, node semantics, or label spaces with a molecular graph. Traditional Graph Neural Networks (GNNs) trained on one domain fail completely when dropped into another.

A powerful realization is reshaping this landscape: while node features change from dataset to dataset, the underlying topological grammar—the fundamental rules of connectivity, motifs, and structures—remains invariant across domains. This insight underpins cross-domain structural transfer, an approach that enables graph foundation models to generalize to entirely unseen networks by learning a universal language of topology.

Structural Invariance as a Shared Vocabulary

How can a model represent a graph without relying on its domain-specific attributes? As argued by Yao Ma in a recent position paper, general-purpose graph models must shift their focus toward a shared vocabulary of structural invariances, such as homophily and triadic closure. If we treat local graph neighborhoods as the "words" of this universal language, we can map completely disparate networks into a unified representation space.

To build intuition, consider how message-passing GNNs operate. A node updates its state by aggregating information from its neighbors, effectively tracing out a local computation tree. Zehong Wang formalizes this by quantizing these local computation trees into discrete tokens, creating a transferable "tree vocabulary" that bridges the gap between graphs with entirely different feature spaces. Similarly, Haonan Yuan proposes decomposing local ego-graphs into structural roots (the invariant motifs) and domain-specific affixes (the local attributes).

By stripping away semantic surface noise, models isolate pure structure. To enforce this feature-agnosticism during training, Fabrizio Frasca et al. introduce Feature-Structuralization, a method that explicitly encodes raw node features into graph topology. This forces the network to learn strictly from shape and connectivity, ensuring that the pre-trained representations remain robust to downstream feature shifts.

Pre-training the Topological Grammar

To master this universal vocabulary, models are pre-trained on massive, multi-domain graph corpora using self-supervised objectives. Instead of relying on task-specific labels, the models learn from the graph's intrinsic properties. For example, Lianghao Xia et al. utilize a self-supervised link prediction objective to train a Graph Mixture-of-Experts architecture, demonstrating that scaling laws apply to graphs just as they do to language when trained on sufficiently diverse topologies.

Crucially, standard pre-training strategies can be counterproductive if they destroy structural identity. Traditional graph contrastive learning relies on random augmentations, such as dropping edges or nodes, which can inadvertently erase the very motifs that define a graph's class. To resolve this, Sihang Li et al. develop a rationale-aware contrastive framework that automatically identifies and preserves invariant structural features. Along the same lines, Xueyi Liu et al. abandon synthetic augmentations altogether, choosing instead to sample real, structurally similar graphs from the dataset to serve as positive pairs.

When models are trained to find these invariant cores across a diverse corpus, they uncover cross-domain synergies. As shown by Hao Liu et al., jointly training on multiple graph domains simultaneously allows a model to capture macro-structural patterns that transcend specific applications—meaning that learning the connectivity rules of a citation network can directly improve predictions on a biological or social network.

Downstream Adaptation via Graph Prompting

Once a model understands this universal structural grammar, how do we apply it to a new domain without destroying its learned weights? The answer lies in graph prompting and in-context learning (ICL).

Instead of fine-tuning the entire network—which can lead to overfitting or catastrophic forgetting—researchers reformulate downstream tasks to match the pre-trained structural space. Qian Huang et al. introduce the first graph ICL framework, PRODIGY, which translates classification tasks into "prompt graphs" that connect few-shot examples directly to queries. This allows the model to reason through a new task entirely in-context, using its structural familiarity to map relationships without updating its parameters.

[Pre-trained GNN Backbone (Frozen)]
               ▲
               │  (Processes universal structural grammar)
       [Prompt Graph]
        ├── Few-Shot Example Nodes
        ├── Target Query Node
        └── Virtual Prompt Edges (Unifying the task format)

For knowledge graphs, Yuanning Cui utilizes a similar prompt-graph paradigm to navigate entirely unseen entities and relations zero-shot. By unifying node, edge, and graph-level tasks into a single graph-prompting format, as demonstrated by Xiangguo Sun et al. and Xixi Wu et al., the core structural parameters of the model can remain frozen while still achieving high performance across completely distinct target domains.

Why It Works: Geometric and Spectral Stability

Why should a structural pattern learned on one network generalize to a completely different one? The deeper intuition relies on the continuous geometric and mathematical properties of graph convolutions.

As proven by Nicolas Keriven et al., as random graphs grow large, the outputs of a GNN converge to stable, continuous operators acting on an underlying geometric space (or manifold). Because graph convolutions essentially function as filters on these continuous shapes, they are fundamentally stable to geometric deformations and non-isomorphic variations. If a source graph and a target graph share similar large-scale continuous properties, the filters learned during pre-training transfer seamlessly.

Furthermore, domain-independent topological signatures—such as Laplacian eigenvectors—provide an analytical bridge between heterogeneous datasets. As shown by Qin Zhang et al., aligning these spectral signatures with semantic features allows models to capture pure structural dynamics. The physics of relational connectivity is mathematically universal; by shifting the focus from what a node represents to how it is structurally embedded, cross-domain transfer unlocks true foundation models for graph-structured data.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 232 papers
  2. Version 1 · 2026-10-04 · 349 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_cross_domain_structural_transfer_00f39ebb,
  title = {cross-domain structural transfer},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-domain-structural-transfer/rdr_00f39ebb8bb240059c19c459cf7ebaf0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_cross_domain_structural_transfer_00f39ebb_v2,
  title = {cross-domain structural transfer},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-domain-structural-transfer/rdr_00f39ebb8bb240059c19c459cf7ebaf0/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake