Fine-Grained Path Augmentation (FGPA)

• FGPA is a framework for graph gradual domain adaptation that constructs intermediate graphs by interpolating source and target domains using the fused Gromov–Wasserstein distance.
• It employs a T-stage self-training regime with confidence weighting to mitigate error accumulation along the optimal geodesic path.
• FGPA integrates with existing graph adaptation methods and demonstrates empirical gains on real-world datasets such as Airport, Citation, and Social graphs.

Fine-Grained Path Augmentation (FGPA) is a framework for graph gradual domain adaptation (GDA), specifically designed to address the challenge of large, non-independent-and-identically-distributed (non-IID) shifts between graph domains. FGPA constructs an optimal sequence of intermediate graphs connecting source and target domains via a geodesic under the Fused Gromov–Wasserstein (FGW) distance and augments self-training with confidence weighting along this interpolated trajectory. This methodology is model-agnostic and compatible with standard graph domain adaptation (graph DA) penalties, enabling robust adaptation in settings where prior approaches assuming mild shifts or given paths are inadequate (2505.12709).

1. Fused Gromov–Wasserstein Distance in Graph Domain Adaptation

FGPA leverages the FGW distance for measuring graph discrepancy. Given attributed graphs $G_0 = (V_0, A_0, X_0)$ and $G_1 = (V_1, A_1, X_1)$, with node-weight histograms $\mu_0 \in \Delta_{|V_0|}$, $\mu_1 \in \Delta_{|V_1|}$, cross-graph feature cost $M(i,j) = \|X_0(i) - X_1(j)\|_{\mathcal{X}}^q$, and adjacency matrices $C_0 = A_0$, $C_1 = A_1$, the FGW distance of order $q$ with trade-off $\alpha$ is: $$d_{\mathrm{FGW};q,\alpha}(G_0, G_1) = \min_{S\in\Pi(\mu_0,\mu_1)} \Bigl[ (1-\alpha)\sum_{i,j} M(i, j) S(i, j) + \alpha \sum_{i,i',j,j'} |C_0(i,i') - C_1(j,j')|^q S(i,j)S(i',j') \Bigr]^{1/q}$$ This formulation unifies node attribute and topological discrepancies and serves as the foundation for quantifying domain shift in the adaptation process. In practical settings, $G_1 = (V_1, A_1, X_1)$0 and $G_1 = (V_1, A_1, X_1)$1 are typical choices (2505.12709).

2. Theoretical Framework and Error Bound in Gradual Adaptation

FGPA follows a $G_1 = (V_1, A_1, X_1)$2-stage self-training regime, denoted by a sequence $G_1 = (V_1, A_1, X_1)$3, where stage $G_1 = (V_1, A_1, X_1)$4 involves adapting model $G_1 = (V_1, A_1, X_1)$5 to $G_1 = (V_1, A_1, X_1)$6 using pseudo-labels. Theoretical analysis (Theorem 3.2) shows that, under standard Lipschitz/Hölder conditions on the per-node loss and GNN architecture, the accumulated error on the target domain satisfies: $G_1 = (V_1, A_1, X_1)$7 where $G_1 = (V_1, A_1, X_1)$8 is the empirical risk, $G_1 = (V_1, A_1, X_1)$9 and $\mu_0 \in \Delta_{|V_0|}$0 encapsulate model smoothness constants, and $\mu_0 \in \Delta_{|V_0|}$1 bounds self-training error. This decomposition clarifies the trade-off between the number of adaptation stages and both self-training accumulation and the total geodesic length traversed in FGW metric (2505.12709).

3. Construction and Properties of the FGW Geodesic Path

The optimal adaptation path is characterized as the FGW geodesic between $\mu_0 \in \Delta_{|V_0|}$2 and $\mu_0 \in \Delta_{|V_0|}$3. By Jensen’s inequality and the metric properties of FGW, the cumulative discrepancy $\mu_0 \in \Delta_{|V_0|}$4 is minimized when each $\mu_0 \in \Delta_{|V_0|}$5 lies at $\mu_0 \in \Delta_{|V_0|}$6 along the interpolating curve: $\mu_0 \in \Delta_{|V_0|}$7 for $\mu_0 \in \Delta_{|V_0|}$8, where $\mu_0 \in \Delta_{|V_0|}$9 are re-arranged adjacency/features in the product node space via optimal low-rank transformations derived from the FGW optimal coupling. Algorithmically, low-rank optimal transport with Dykstra updates is used to find the coupling, after which intermediate graphs are generated by linear interpolation (2505.12709).

4. Self-Training along the FGW Geodesic with Confidence Modulation

FGPA performs adaptation via self-training on each intermediate graph. At each stage, predictions from $\mu_1 \in \Delta_{|V_1|}$0 form pseudo-labels for $\mu_1 \in \Delta_{|V_1|}$1. Confidence scores are calculated for each node using normalized entropy: $\mu_1 \in \Delta_{|V_1|}$2 where $\mu_1 \in \Delta_{|V_1|}$3 denotes Shannon entropy. These confidences are employed to down-weight noisy predictions in the stage-wise supervised loss: $\mu_1 \in \Delta_{|V_1|}$4 This denoising mechanism counteracts self-training noise accumulation inherent in GDA (2505.12709).

5. Integration with Existing Graph DA Workflows

FGPA is orthogonal to specific domain adaptation loss choices and can be used in conjunction with existing graph DA techniques such as MMD, CORAL, AdaGCN, GRADE, StruRW, adversarial or spectral regularizations. The total loss at each stage can incorporate an arbitrary graph DA penalty: $\mu_1 \in \Delta_{|V_1|}$5 FGPA thus serves as a path augmentation module, enhancing any base DA method with fine-grained geodesic trajectory information, rather than altering the adaptation loss structure itself. This modularity is a distinguishing feature of the approach (2505.12709).

6. Empirical Evaluation and Results

FGPA has been evaluated on diverse datasets, including real-world graphs—Airport (USA⇄Europe⇄Brazil), Citation (ACM⇄DBLP), Social (Blog1⇄Blog2)—and synthetic contextual SBMs with controlled shifts. Backbone architectures include 2–3 layer GCNs or APPNP with hidden dimensions 8–16. Standard adaptation baselines (ERM, MMD, CORAL, AdaGCN, GRADE, StruRW) serve as comparators.

Key experimental settings and results:

Dataset	Average Gain (pp)	Max Gain (pp)	Synthetic Gain (pp)
Airport	+6.8	+26.3	—
Social	+3.6	—	—
Citation	+3.4	—	—
CSBM (synthetic)	—	—	+36.5

• Geodesic path construction: $\mu_1 \in \Delta_{|V_1|}$6 intermediate graphs, $\mu_1 \in \Delta_{|V_1|}$7, $\mu_1 \in \Delta_{|V_1|}$8, low-rank OT of rank $\mu_1 \in \Delta_{|V_1|}$9.
• Training: Adam optimizer, learning rate $M(i,j) = \|X_0(i) - X_1(j)\|_{\mathcal{X}}^q$0, 1000 epochs, 5 random seeds.

FGPA improved node classification accuracy over direct adaptation (one-shot baseline) in over 90% of real-world domain adaptation tasks, with worst-case degradation under 2.6 percentage points in cases of mild shift (2505.12709).

7. Practical Implications and Future Directions

FGPA establishes a principled framework for generating and exploiting intermediate domains in the space of attributed graphs under severe non-IID shifts. By leveraging the optimality properties of the FGW geodesic and the flexibility of confidence-modulated self-training, it addresses robustness issues in graph DA scenarios where traditional metrics and heuristics fail. A plausible implication is the extension of this pipeline to other structured data modalities that admit Gromov–Wasserstein-type interpolations. Further, the plug-and-play compatibility with arbitrary domain adaptation losses supports a modular ecosystem for future graph adaptation research.