Graphon-Informed Augmentations for Graph Learning

• Graphon-Informed Augmentations are principled data augmentation techniques that generate synthetic graphs from estimated graphons to capture global structural patterns.
• They leverage advanced graphon estimation methods, including Gromov–Wasserstein barycenters and INR-based moment matching, to sample class- or cluster-specific graphs.
• Empirical results demonstrate that GIAs achieve consistent improvements in supervised and self-supervised learning compared to traditional heuristic augmentations.

Graphon-Informed Augmentations (GIAs) are a family of principled data augmentation strategies for graph-based machine learning, grounded in the theory of graph limits and graphon estimation. Unlike standard node- or edge-level perturbations, GIAs construct class- or cluster-specific generative models—graphons—that capture the global structural patterns of observed graphs. Synthetic graphs are then sampled from these graphons or from graphons estimated to interpolate or adapt real graph statistics, resulting in augmented datasets that reflect structurally coherent, data-driven variations. GIAs have been applied in supervised learning, self-supervised contrastive learning, and mixup-style graph classification, demonstrating consistent improvements over heuristic augmentation baselines.

1. Definition and Theoretical Basis

GIAs utilize graphons, which are symmetric measurable functions $W : [0,1]^2 \to [0,1]$, as probabilistic generative models for random graphs. A graphon defines a distribution over graphs of any size via latent position sampling: for $N$ nodes, draw latent coordinates $(x_1, ..., x_N) \sim \mathrm{Uniform}[0,1]^N$ and instantiate each edge independently with probability $W(x_i, x_j)$. GIAs operate at “the level of the generative mechanism,” rather than directly on observed discrete graphs (Ponti, 2024, Azizpour et al., 6 Jun 2025, Ramezanpour et al., 4 Jun 2025).

In supervised scenarios, each class label corresponds to a class-conditional graphon estimate $\widehat{W}_i$, from which new labeled graphs are sampled to enrich or balance the training set. In the context of self-supervised contrastive learning, GIAs sample plausible structural perturbations by resampling edges according to an estimated graphon conditional on latent node coordinates (Azizpour et al., 6 Jun 2025). For class-mixing or interpolation tasks, GIAs can operate in the space of graph moments (“moment space”), ensuring motif counts or subgraph statistics interpolate along convex paths between classes (Ramezanpour et al., 4 Jun 2025).

2. Graphon Estimation Methods

Estimation of the generative graphon from observed graphs is central to GIAs. Algorithms exploit both classical graph limit theory and modern neural architectures:

• Gromov–Wasserstein Barycenter Estimation: Given $M$ unaligned adjacency matrices, the estimator solves

$$\min_W \frac{1}{M} \sum_{m=1}^M d_{\mathrm{gw},2}^2(A_m, W)$$

where $d_{\mathrm{gw},2}$ is the squared 2-order Gromov–Wasserstein distance, a permutation-invariant non-Euclidean metric optimized via alternating minimization and optionally entropic smoothing (Ponti, 2024).

• SIGL (Scalable Implicit Graphon Learning): This method jointly estimates node latent coordinates using a GNN and learns an implicit neural representation (INR) of the graphon, aligning reconstructed adjacency matrices to observed graphs. It proceeds via (i) latent variable estimation, (ii) blockwise histogram averaging on sorted adjacency matrices, and (iii) regression on sampled INR points to fit the continuous graphon (Azizpour et al., 6 Jun 2025).
• Moment Matching with INRs: This approach forgoes latent variables and instead trains an INR $f_\theta$ to match empirical motif frequencies (“moments”) from input graphs, minimizing

$$\mathcal{L}(\theta) = \sum_{k=1}^K w_k (m_k^{\mathrm{target}} - \hat{t}'(F_k, f_\theta))^2$$

where $m_k^{\mathrm{target}}$ is a (possibly mixed) target motif vector and $\hat{t}'(F_k, f_\theta)$ is the INR’s predicted motif density via Monte Carlo integration over latent positions (Ramezanpour et al., 4 Jun 2025).

3. GIA Algorithms for Data Augmentation

The GIAs paradigm comprises several workflows, unified by sampling from estimated or constructed graphons:

• Class-Conditional GIAs: For each class, estimate $\widehat{W}_i$ and sample new graphs as follows:
  1. For graph size $N$, sample $x_1, ..., x_N \sim \mathrm{Uniform}[0,1]$
  2. For $i<j$, sample edge $$A_{ij} \sim \mathrm{Bernoulli}(\widehat{W}_i(x_i,x_j))$$
  3. Augment the labeled dataset with these synthetic samples (Ponti, 2024, Ramezanpour et al., 4 Jun 2025).

• Graphon-Informed Edge Perturbation: Given a single graph and an estimated graphon (with node latent coordinates), stochastically resample a fixed proportion $r\%$ of possible edges using the graphon’s generative probabilities; untouched edges are retained from the original graph (Azizpour et al., 6 Jun 2025).

• MomentMixup Augmentation: Compute empirical motif vectors from two graphs or classes, interpolate in moment space,

$m^{\mathrm{mix}} = \lambda m_i + (1 - \lambda) m_j$

and find $\widehat{W}_{\theta^*}$ whose motif densities approximate $m^{\mathrm{mix}}$; sample augmented graphs from $\widehat{W}_{\theta^*}$ (Ramezanpour et al., 4 Jun 2025).

Pseudocode for MomentMixup:

Given graphs G_i (class i) and G_j (class j): 1. Compute moments m_i and m_j from N_sample graphs in each class 2. Draw λ ~ Beta(α, α); set m_mix = λ*m_i + (1-λ)*m_j 3. Train INR f_θ (graphon estimator) to match moments to m_mix 4. For N_aug times: - Sample η_1,...,η_n ~ Uniform[0,1] - For i<j, draw edges by f_θ(η_i, η_j) - Output graph with interpolated label y_mix = λ*y_i + (1-λ)*y_j

4. Applications in Learning Frameworks

GIAs have been adapted to various learning settings:

• Supervised Graph Classification: Augmenting class datasets with samples from estimated class graphons yields accuracy gains across synthetic (LFR) and real-world (IMDB, ENZYMES) benchmarks. Gromov–Wasserstein barycenter estimators generally outperform stochastic block model or low-rank baselines (Ponti, 2024).
• Self-Supervised Contrastive Learning: In the MGCL framework, GIAs provide data-driven, semantically faithful augmentations by edge resampling via graphon-conditioned probabilities. For graph-level tasks, clustering is used to fit multiple graphons corresponding to latent structure groups, and only structurally dissimilar clusters are contrasted, leading to improved representation quality (Azizpour et al., 6 Jun 2025).
• Graph Mixup: MomentMixup enables mixup-style augmentation by interpolating in motif (“moment”) space and inverting to a graphon representation, which is used to generate graphs whose subgraph statistics lie along the intended interpolation path between classes (Ramezanpour et al., 4 Jun 2025).

5. Theoretical Guarantees

Several theoretical properties of GIAs have been established:

• Consistency and Distance Bounds: The Gromov–Wasserstein barycenter provides consistent graphon estimation in cut-metric under oracle alignment; minimizing the GW distance yields better approximations without requiring node correspondence (Ponti, 2024).
• Motif–Cut Distance Lemma: If two graphons $W_1, W_2$ match on all motif densities for motifs up to size $k$ within $\delta_M$, their cut distance is at most $O(1/\sqrt{\log k})$ for $\delta_M = 3^{-k^2}$. This supports moment-based GIAs in controlling the difference between generated and target class graphons (Ramezanpour et al., 4 Jun 2025).
• Empirical Convergence: When INR-based moment matching achieves small motif error and empirical moments concentrate due to large sample size, the estimated graphon is provably close to the true generative graphon in cut distance, with high probability (Ramezanpour et al., 4 Jun 2025).

6. Empirical Results and Comparative Effectiveness

GIAs exhibit consistent empirical advantages over standard augmentations. Key findings include:

Dataset	No Augment.	G-Mixup (USVT)	G-Mixup (SIGL)	MomentMixup
IMDB-Binary	71.6 ± 3.5	71.9 ± 3.0	74.0 ± 2.6	74.3 ± 2.7
IMDB-Multi	48.8 ± 2.8	50.5 ± 1.5	50.7 ± 1.4	50.9 ± 1.9
REDD-Binary	91.8 ± 1.1	91.3 ± 1.5	92.3 ± 1.4	91.8 ± 1.2
AIDS	98.0 ± 1.2	97.8 ± 0.9	97.3 ± 1.0	98.5 ± 0.6

On LFR, IMDB, and ENZYMES, augmenting with as few as 1% synthetic samples leads to measurable accuracy improvements. GW-barycenter-based GIAs are consistently among the best-performing methods, with MomentMixup achieving the top results on selective benchmarks (Ponti, 2024, Ramezanpour et al., 4 Jun 2025). In contrastive learning, GIAs yield semantically faithful augmentations and reduce false negatives in InfoNCE objectives by leveraging latent generative structure (Azizpour et al., 6 Jun 2025).

7. Methodological Considerations and Extensions

Key distinctions and extension avenues for GIAs include:

• Graphon Domain vs. Heuristic Augmentations: GIAs employ data-adaptive, model-driven sampling schemes, assigning each edge a resampling probability dictated by the graphon, in contrast to uniform or random perturbations that may disrupt crucial substructures (Azizpour et al., 6 Jun 2025).
• Interpolation Pathways: MomentMixup corrects a limitation of naive “graphon mixing” by ensuring that motif densities interpolate linearly, preserving motif-level structure means between classes (Ramezanpour et al., 4 Jun 2025).
• Potential Extensions: Richer collections of motif moments, attribute-conditioned moments, time-varying signatures for dynamic networks, nonconvex (e.g., geodesic) moment interpolation, and integration of GIAs in contrastive regimes are proposed future directions (Ramezanpour et al., 4 Jun 2025).

In summary, Graphon-Informed Augmentations constitute a rigorous, theory-backed approach to graph data augmentation, harnessing advances in graphon estimation and moment matching to synthesize structure-aligned, semantically coherent training data, with demonstrated empirical and theoretical superiority over heuristic perturbation baselines in diverse graph learning scenarios (Ponti, 2024, Azizpour et al., 6 Jun 2025, Ramezanpour et al., 4 Jun 2025).