SA$^{2}$GFM: Enhancing Robust Graph Foundation Models with Structure-Aware Semantic Augmentation

Junhua Shi, Qingyun Sun, Haonan Yuan, Xingcheng Fu

TL;DR

SA²GFM is a new framework designed to make Graph Foundation Models (GFMs) significantly more robust against noise and adversarial attacks. By injecting hierarchical structural information via textual prompts and using a "null expert" to filter out irrelevant cross-domain knowledge, it achieves state-of-the-art performance even when graphs are heavily corrupted.

What Problem Does This Paper Address?

Graph Foundation Models (GFMs) aim to be the "GPT of graphs," learning general representations across multiple domains that can be easily adapted to specific tasks. However, they face two major hurdles:

  1. Fragility: Most GFMs use shallow Graph Neural Network (GNN) backbones that are theoretically limited by the 1-Weisfeiler-Lehman (1-WL) test. This means they struggle to distinguish complex structural patterns and are highly vulnerable to random noise or targeted adversarial attacks (like adding or removing specific edges).
  2. Negative Transfer: When adapting a model from one domain (e.g., citation networks) to another (e.g., product co-purchase graphs), the differences in structure can actually degrade performance if the model tries to force-fit irrelevant knowledge.

Methods

The authors propose SA²GFM (Structure-Aware Semantic Augmentation GFM), which tackles these issues through three main components:

  • Structure-Aware Semantic Augmentation: The model uses graph structural entropy to build a hierarchical "encoding tree." This tree is converted into natural language prompts (e.g., "Node 1 belongs to cluster A which contains 2 nodes..."). These prompts are processed by a language model (BERT) to enrich the raw node features with high-level structural context.
  • Self-Supervised Information Bottleneck (SS-IB): To ensure the model doesn't just memorize noise, the authors implement an Information Bottleneck objective. This forces the model to compress redundant input information while maximizing the "signal" needed for downstream tasks.
  • Expert Adaptive Routing with a Null Expert: During adaptation, the model uses a Mixture-of-Experts (MoE) architecture. Crucially, it includes a null expert—a shallow GCN trained only on the target graph. If the pre-trained experts from other domains are a poor fit, the router assigns more weight to the null expert, effectively "rejecting" harmful knowledge transfer.
Figure 2: Overview of the SA2GFM framework. The framework first pre-trains an ensemble of expert encoders on multiple source domains using Structure-Aware Semantic Augmentation and an Information Bottleneck objective. During fine-tuning, a novel Expert Adaptive Routing mechanism selects knowledge to mitigate negative transfer, while a Hierarchical Structure Optimization module refines the target graph’s topology before final prediction.

Datasets

The model was evaluated across seven benchmark datasets spanning three distinct domains:

  • Citation Domain: Cora, CiteSeer, PubMed, and ogbn-arXiv.
  • Products Domain: ogbn-Tech and ogbn-Home.
  • Web Page Domain: Wiki-CS.

Evaluation focused on two challenging scenarios: Cross-Dataset (same domain, different dataset) and Cross-Domain (entirely different fields).

Key Insights and Ideas

  • Text as a Structural Bridge: One of the most creative ideas is using LLMs to bridge the gap between structure and semantics. By describing a node's structural role in words, the model leverages the rich semantic space of pre-trained language models to understand graph topology.
  • The Power of "No": The "null expert" is a simple but powerful solution to the negative transfer problem. It acknowledges that sometimes, the best thing a foundation model can do is realize it doesn't know enough about a new domain and rely on local information instead.
  • Hierarchical Refinement: Unlike many models that treat graph structure as a flat set of edges, SA²GFM optimizes the graph at two levels during fine-tuning: intra-cluster (local fidelity) and inter-cluster (global signal propagation).
Figure 3: Construct code tree by minimizing structure en- tropy to generate structure aware prompt text.

Key Results

SA²GFM was compared against 9 state-of-the-art baselines and consistently came out on top:

  • Superior Robustness: On average, it achieved a +5.9% accuracy improvement in node classification and +2.4% in graph classification over the best-performing runner-up.
  • Stability Under Attack: As the intensity of noise or adversarial perturbations increased, SA²GFM's performance declined much more slowly than its competitors. In severe noise settings ($\lambda = 0.8$), it maintained a significant lead.
  • Cross-Domain Success: In challenging cross-domain transfers, it achieved a +5.1% average gain over MDGFM, the strongest existing robust baseline.
Performance comparison under varying degrees of random noise and targeted attacks.

Limitations and Future Work

While robust, the authors note several areas for improvement:

  • Complex Attacks: The current evaluation focuses on standard perturbations. Future work could test resilience against more sophisticated, multi-step topological attacks.
  • Graph Diversity: The framework has primarily been validated on homogeneous, static graphs. Extending SA²GFM to heterogeneous (multiple node/edge types) or dynamic (time-varying) graphs is a logical next step.
  • Computational Cost: While the authors argue the structure optimization is efficient, the use of BERT for feature augmentation adds an inference overhead that might be challenging for extremely large-scale real-time applications.

Summary updated

Abstract

We present Graph Foundation Models (GFMs) which have made significant progress in various tasks, but their robustness against domain noise, structural perturbations, and adversarial attacks remains underexplored. A key limitation is the insufficient modeling of hierarchical structural semantics, which are crucial for generalization. In this paper, we propose SA$^{2}$GFM, a robust GFM framework that improves domain-adaptive representations through Structure-Aware Semantic Augmentation. First, we encode hierarchical structural priors by transforming entropy-based encoding trees into structure-aware textual prompts for feature augmentation. The enhanced inputs are processed by a self-supervised Information Bottleneck mechanism that distills robust, transferable representations via structure-guided compression. To address negative transfer in cross-domain adaptation, we introduce an expert adaptive routing mechanism, combining a mixture-of-experts architecture with a null expert design. For efficient downstream adaptation, we propose a fine-tuning module that optimizes hierarchical structures through joint intra- and inter-community structure learning. Extensive experiments demonstrate that SA$^{2}$GFM outperforms 9 state-of-the-art baselines in terms of effectiveness and robustness against random noise and adversarial perturbations for node and graph classification.
BibTeX
@inproceedings{shi2026sa,
  author = {Junhua Shi and Qingyun Sun and Haonan Yuan and Xingcheng Fu},
  title = {{SA$^\{2\}$GFM: Enhancing Robust Graph Foundation Models with Structure-Aware Semantic Augmentation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  doi = {10.1609/aaai.v40i18.38602},
}
Figure