Pre-trained Graph Condensation (PreGC)

• Pre-trained Graph Condensation (PreGC) is a paradigm that condenses large graphs into compact, reusable representations to support model-agnostic pre-training.
• It employs optimal transport and diffusion augmentation to align semantic and structural graph features for robust transfer across tasks.
• PreGC overcomes conventional condensation limitations by decoupling backbone biases and enabling reusable, task-independent synthetic graphs.

Pre-trained Graph Condensation (PreGC) denotes a graph condensation paradigm in which a large graph is distilled into a small synthetic graph intended to remain useful beyond a single dataset, task, or relay architecture. In contrast to conventional graph condensation, which commonly optimizes a condensed graph for a fixed supervised objective and a fixed GNN, PreGC emphasizes task-independent or transfer-oriented condensation, backbone generalization, and reusable condensed-data pre-training protocols. In the current literature, this direction is made explicit by "Towards Pre-trained Graph Condensation via Optimal Transport" (Yan et al., 18 Sep 2025), is anticipated by transferable condensation methods such as TGCC (Du et al., 29 Jan 2026), and is contextualized by the benchmarking and survey literature on graph condensation (Sun et al., 2024, Gao et al., 2024).

1. Emergence within the graph condensation literature

Graph condensation (GC) is defined as synthesizing a compact yet highly representative graph that enables GNNs trained on it to achieve performance comparable to those trained on the original large graph (Gao et al., 2024). The survey literature organizes GC methods around effectiveness, generalization, fairness, efficiency, and robustness, and describes generalized GC as the branch concerned with adaptability across GNN architectures and downstream tasks rather than only task-specific accuracy (Gao et al., 2024).

The term Pre-trained Graph Condensation is not presented by the survey as an already established subfield; instead, the survey states that it does not explicitly name “Pre-trained Graph Condensation (PreGC)” as an established area, while also noting that its principles are closely related to generalized graph condensation (Gao et al., 2024). This places PreGC historically as a specialization of the broader effort to make condensed graphs model-agnostic and task-agnostic.

GC-Bench sharpens this picture by making pre-training an explicit evaluation scenario. It introduces a comprehensive and open benchmark that evaluates 12 state-of-the-art GC algorithms across node-level and graph-level tasks on 12 diverse datasets, and it states that its framework and protocols facilitate the training and subsequent fine-tuned evaluation of condensed datasets, including pre-training support (Sun et al., 2024). GC-Bench also notes that none of the original benchmarked methods are strictly “pre-trained condensation” algorithms by design, even though the benchmark supports such evaluation. This suggests that PreGC first emerged as an evaluation need before it became a method family in its own right.

2. Problem formulation and the limits of conventional GC

The generic GC objective is commonly written as

$$\mathcal{S} = \arg\min_{\mathcal{S}} \mathcal{L}_{cond}\left(f_{\theta}(\mathcal{T}), f_{\theta}(\mathcal{S})\right),$$

where $\mathcal{T}$ is the original graph, $\mathcal{S}$ is the condensed graph, $f_\theta$ is a GNN or relay model, and $\mathcal{L}_{cond}$ is the condensation objective (Gao et al., 2024). The survey identifies four major optimization families behind this objective: gradient matching, trajectory matching, kernel ridge regression, and distribution matching (Gao et al., 2024). In practice, many such methods are tied to a particular supervision signal and a particular relay architecture.

GC-Bench provides the clearest empirical diagnosis of why this is insufficient for PreGC. It reports that condensed graphs optimized for one task, such as node classification, perform poorly when transferred to other tasks such as link prediction, anomaly detection, or node clustering; performance is always less than or equal to original data and is sometimes much worse (Sun et al., 2024). The same benchmark reports that condensed graphs transfer reasonably only between similar or simple GNNs such as SGC and GCN, but fail when transferred to models with different inductive biases, such as Graph Transformer; extensive gradient or trajectory matching can overfit the condensed dataset to the backbone, and condensed data effectively encodes the “pre-trained” backbone bias (Sun et al., 2024).

"Towards Pre-trained Graph Condensation via Optimal Transport" (Yan et al., 18 Sep 2025) reformulates this issue from the perspective of GNN optimization consistency. It derives a generalized objective,

$$\tilde{\mathcal{G}}^* = \arg\min_{\tilde{\mathcal{G}}} \left\{ \Delta_\mathcal{Z}(\mathbf{Z}, \tilde{\mathbf{Z}}) + \xi \Delta_\mathcal{Y}(\mathcal{G},\tilde{\mathcal{G}}) \right\},$$

where $\Delta_\mathcal{Z}$ is a representation-level discrepancy and $\Delta_\mathcal{Y}$ is a semantic-level discrepancy (Yan et al., 18 Sep 2025). The central change is that $\Delta_\mathcal{Y}$ is redefined to be task-agnostic through transport-based semantic alignment rather than label matching. This is the formal step that distinguishes explicit PreGC from ordinary supervised GC.

3. Task-independent PreGC via optimal transport

The optimal-transport formulation is the first method in the supplied literature that names itself directly as PreGC and characterizes its goal as transcending task- and architecture-dependent graph condensation (Yan et al., 18 Sep 2025). Its methodology has three coupled components: hybrid-interval graph diffusion augmentation, optimal transport plan matching, and a traceable semantic harmonizer.

The diffusion component is designed to improve architecture-agnosticity. It starts from the graph heat equation,

$$\frac{d\mathbf{Z}_t}{dt} = -\mathbf{L}\mathbf{Z}_t,$$

and discretizes it by explicit Euler so that

$\mathcal{T}$0

with terminal state

$\mathcal{T}$1

At each training epoch, $\mathcal{T}$2 is sampled from a uniform distribution $\mathcal{T}$3, and the resulting diffusion states are aligned between source and condensed graphs (Yan et al., 18 Sep 2025). The paper states that this hybrid-interval graph diffusion augmentation suppresses weak generalization on particular architectures by enhancing the uncertainty of node states.

The transport component matches semantics in both graph space and representation space. At the graph level, the method uses the Fused Gromov-Wasserstein Distance to couple node features and structure:

$\mathcal{T}$4

At the representation level, it computes a Wasserstein plan on diffusion-augmented embeddings,

$\mathcal{T}$5

These plans are then matched by

$\mathcal{T}$6

and combined with the diffusion alignment cost in the total loss (Yan et al., 18 Sep 2025).

The semantic harmonizer is the mechanism that turns a pre-trained condensed graph into a reusable downstream object. The optimized representation transport plan $\mathcal{T}$7 gives a traceable association from source nodes to condensed nodes. After discretization into an assignment matrix $\mathcal{T}$8, labels from any source training subset $\mathcal{T}$9 can be transferred by

$\mathcal{S}$0

The paper presents this as dynamic, flexible label allocation without re-condensation, and states that extensive experiments verify the superiority and versatility of PreGC, demonstrating its task-independent nature and seamless compatibility with arbitrary GNNs (Yan et al., 18 Sep 2025). A plausible implication is that, within the supplied corpus, this is the clearest methodological realization of PreGC as a reusable pre-training object rather than merely a transfer evaluation protocol.

4. Transferable condensation from the causal perspective

TGCC, proposed in "Transferable Graph Condensation from the Causal Perspective" (Du et al., 29 Jan 2026), is framed as a transferable graph dataset condensation method rather than a task-independent unsupervised PreGC method, but its design is explicitly presented as suitable for a pre-trained condensation stage. The paper identifies the main failure mode of traditional GC as the encoding of correlations specific to the original dataset and task, which leads to poor cross-task and cross-domain transfer (Du et al., 29 Jan 2026).

Its architecture contains three modules. Causal Invariant Feature Extraction (CIFE) perturbs only the high-frequency spectral components of the adjacency matrix to generate an intervened graph and enforces dimension-wise consistency between the representations of the original and intervened graphs, while discouraging spurious dependencies through HSIC or covariance penalties. The overall causal loss is written as

$\mathcal{S}$1

The paper states that Theorem 1 shows optimizing this loss upper bounds the difference between corresponding spectral amplitudes of the low-frequency parts of the original and intervened graphs (Du et al., 29 Jan 2026).

Graph Contrastive Condensation (GCC) extends gradient matching so that the synthetic graph matches learning behavior on both the original graph and its causally intervened counterpart, using a bi-level objective over training trajectories. Spectral-domain Enhanced Contrastive Learning (ECL) then injects causal-invariant information into the condensed graph through spectral-domain negatives and an InfoNCE loss. The full loss is

$\mathcal{S}$2

The experimental results are explicitly transfer-oriented. On five public datasets and the novel FinReport dataset, TGCC achieves up to a 13.41\% improvement in cross-task and cross-domain complex scenarios compared to existing methods, and it achieves state-of-the-art performance on 5 out of 6 datasets in the single dataset and task scenario (Du et al., 29 Jan 2026). The detailed results further state that all three modules are necessary for best transfer performance, that TGCC is 3x faster than some recent state-of-the-art condensation approaches, and that it generalizes across GCN, SAGE, SGC, APPNP, Cheby (Du et al., 29 Jan 2026). The paper also states that TGCC’s condensation framework is plug-and-play and can enhance many condensation pipelines as a pre-trained condensation stage (PreGC), producing condensed datasets that can be widely shared and used for various downstream tasks and domains. In this sense, TGCC occupies an intermediate position between transfer-oriented GC and explicit PreGC.

5. Enabling design patterns from efficient, decoupled, and robust GC

Several methods in adjacent parts of the GC literature are not introduced as fully task-independent PreGC algorithms, but they contribute design patterns repeatedly linked to pretraining-style reuse, backbone transfer, or repeated condensation. These patterns are especially visible in decoupling strategies, use of frozen pre-trained modules, and robustness-oriented objectives.

SP-ESGC separates node condensation from graph structure generation and uses heat kernel feature propagation, a hybrid clustering strategy, and a pre-trained edge predictor to synthesize structure (Hu et al., 29 May 2026). The paper states that this decoupled design avoids costly entangled optimization, implements precise GC with significantly high computational efficiency, and generalizes well across diverse GNN architectures (Hu et al., 29 May 2026). This suggests a structure-preserving route to reusable condensed graphs even though the paper is not framed directly as a task-independent PreGC method.

DisCo makes the decoupling more explicit by independently condensing nodes and generating edges in a two-stage GNN-free paradigm (Xiao et al., 2024). Node condensation is performed against a pre-trained node classification model, and edge construction is transferred from a pre-trained link prediction model. The paper reports that this approach is at least 10 times faster than state-of-the-art methods on medium-scale graphs, can scale to Ogbn-papers100M, and is robust across GCN, SAGE, GIN, JKNet, even MLP (Xiao et al., 2024). It also argues that once the node classification and link prediction models are trained on the original large graph, multiple condensed graphs can be created without retraining these teachers, which directly aligns with PreGC-style reuse (Xiao et al., 2024).

SimGC replaces trainable surrogate GNNs with a frozen pre-trained Simple Graph Convolution (SGC) guide and aligns condensed and original graphs from the input layer to the prediction layer through per-class mean and standard deviation statistics, logits, and feature smoothness (Xiao et al., 2024). The paper reports up to 10 times speedup compared with existing graph condensation methods while performing on par with state-of-the-art baselines, and it emphasizes generalization capability across MLP, GCN, SGC, GraphSAGE, GIN, JKNet (Xiao et al., 2024). Its implication for PreGC is parameter-efficiency: a pre-trained, non-trainable guide may suffice for effective condensation.

GroC and its time-efficient variant TimGroC contribute a robustness-oriented pattern. They introduce the Shock Absorber, an adversarial perturbation that selectively attacks weak or underrepresented parts of the synthetic graph during gradient matching, while sharing the backward process in a free training manner (Li et al., 2023). The paper reports about 0.2\% to 2.2\% additional time overhead and nearly 4-fold improvement over general adversarial training, with positive cross-backbone behavior that it explicitly describes as suitable for PreGC settings (Li et al., 2023). This suggests that robustness objectives can serve not only stability but also transferability in condensed graphs.

Method	Design choice	Reported PreGC relevance
SP-ESGC	Decoupled node condensation and pre-trained edge predictor	Generalizes well across diverse GNN architectures
DisCo	GNN-free node condensation and transferred link prediction	Practical for PreGC pipeline and repeated condensation
SimGC	Frozen pre-trained SGC guide	Generalization capability with up to 10 times speedup
GroC / TimGroC	Shock Absorber with free adversarial training	Better transfer across architectures; suitable for PreGC settings

6. Benchmarks, misconceptions, and open directions

The benchmark literature presents a restrained picture of current progress. GC-Bench shows that pre-training on condensed data can improve downstream node clustering relative to naive transfer, but it still lags far behind original data (Sun et al., 2024). It also shows that architecture transfer remains strongly biased toward the backbone used during condensation, and that current condensation methods overfit condensed graphs to a specific backbone or task; task-agnostic pretraining therefore remains unsolved in the benchmark’s analysis (Sun et al., 2024). A common misconception is that any condensed graph that performs well on node classification can automatically serve as a universal pre-training corpus. The benchmark evidence does not support that conclusion.

A second misconception is that PreGC is already a settled, clearly bounded area. The survey explicitly says that it does not explicitly name “Pre-trained Graph Condensation (PreGC)” as an established subfield, even though generalized GC principles closely track what PreGC seeks to achieve (Gao et al., 2024). The literature therefore contains both explicit PreGC methods and a broader penumbra of transfer-oriented or architecture-robust condensation methods.

The principal open directions are also consistent across the benchmark and survey literature. The survey highlights the need for condensation methods for heterophilic, heterogeneous, dynamic, and directed graphs, for task-agnostic condensation / self-supervised PreGC, for model/task-agnostic assessment metrics that evaluate condensed-graph quality without expensive downstream retraining, and for stronger work on robustness and security and explainability (Gao et al., 2024). GC-Bench adds an operational challenge: time, memory, and disk demands remain prohibitive for several gradient- and trajectory-matching methods, especially at larger condensation ratios or larger graphs (Sun et al., 2024). Taken together, these results indicate that PreGC is best understood not as a completed solution, but as a research program aimed at producing compact graphs that are reusable, semantically traceable, architecture-agnostic, and resistant to the task-specific and model-specific biases of earlier GC methods.