---
title: Fused Gromov-Wasserstein Transport
url: https://www.emergentmind.com/topics/fused-gromov-wasserstein-optimal-transport
type: topic
---

# Fused Gromov-Wasserstein Transport

The Fused Gromov-Wasserstein (FGW) optimal transport framework generalizes classical Wasserstein and Gromov-Wasserstein distances to provide a unified metric for comparing structured objects that possess both geometric and attribute information—most notably, graphs and metric-measure spaces with node features. By blending feature-based optimal transport and structural relational comparisons, FGW enables flexible alignment of complex data modalities, supports barycenter computation for clustering and averaging, and drives modern machine learning architectures for graph-based prediction, classification, augmentation, and alignment. Recent developments include unbalanced and partial-matching extensions, scalable algorithms, variance decompositions via LOT embeddings, and formulations for edge attributes and multi-marginal settings.

## 1. Formal Definition and Mathematical Structure

Given two metric-measure spaces $\mathbb{X}=(X,d_X,\mu)$ and $\mathbb{Y}=(Y,d_Y,\nu)$ (or, in the graph context, attributed graphs $(X,C_X,a_X)$ and $(Y,C_Y,a_Y)$), the FGW distance of order $(r,q)$ is defined as
\[
\mathrm{FGW}_{r,L}(\mathbb X,\mathbb Y) :=\min_{\gamma\in\Gamma(\mu,\nu)} \left[ \omega_1\,\langle C,\gamma\rangle \;+\;\omega_2\,\gamma^{\otimes2}\left(L\left(d_X^r,d_Y^r\right)\right) \right]
\]
where:
- $C:X\times Y\to\mathbb{R}_+$ is the feature cost (e.g. $C(x,y)=\|x-y\|^q$),
- $L:\mathbb{R}^2\to\mathbb{R}_+$ is the structural cost (e.g. $L(a,b)=|a-b|^q$),
- $\omega_1, \omega_2\geq0$, $\omega_1+\omega_2=1$ trade-off feature and structure,
- $\gamma$ is a coupling in $\Gamma(\mu, \nu)$ matching marginals $\mu$ and $\nu$,
- $\gamma^{\otimes2}(L(\cdot,\cdot))$ is an expectation of pairwise structural discrepancies over the coupling.

FGW interpolates between classical OT ($\omega_1=1$) and pure GW ($\omega_2=1$) [2502.09934], [1811.02834].

## 2. Variants and Extensions: Unbalanced and Partial FGW

Classical FGW enforces mass equality. To address unbalanced and noisy data, several extensions have recently appeared:

- **Fused Partial Gromov-Wasserstein (FPGW)**: Constraints can be relaxed either via a total-variation penalty (mass deviation) or a direct constraint on the mass transported. Two key forms are equivalent (Thm. 3.1):
  - TV-penalty:
    \[
    \mathrm{FPGW}_{r,L,\lambda}(\mathbb X,\mathbb Y) :=\min_{\gamma\in\mathcal{M}_+(X\times Y)} \left\{ \text{FGW terms} + \lambda|\mu^{\otimes2}-\gamma_1^{\otimes2}|_{TV} + \lambda|\nu^{\otimes2}-\gamma_2^{\otimes2}|_{TV} \right\}
    \]
  - Mass-Constrained:
    \[
    \mathrm{FMPGW}_{r,L,\rho}(\mathbb X,\mathbb Y) :=\min_{\gamma\in \Gamma^\rho_\le(\mu,\nu)} \text{FGW terms}
    \]
    with $\gamma_1\leq\mu, \gamma_2\leq\nu, |\gamma|\geq\rho$.

- **Fused Unbalanced Gromov-Wasserstein (FUGW)**: Relax hard marginal constraints with KL penalties [2206.09398].
  \[
    \mathrm{FUGW}_{\beta,\lambda}(T) = (1-\lambda)\sum_{i,k}d_F(f_i,g_k)^2\,T_{ik} + \lambda\sum_{i,j,k,l}|a_{ij}-b_{kl}|^2\,T_{ik}T_{jl} + \varepsilon \mathrm{KL}(T\mathbf{1}\|\mu) + \varepsilon \mathrm{KL}(T^T\mathbf{1}\|\nu)
  \]
  These forms permit selective matching and robustness to outliers [2502.09934], [2206.09398], [2406.19767], [2509.24382].

## 3. Algorithmic Frameworks: Optimization and Scalability

FGW and its variants involve nonconvex quadratic objectives solved predominantly by:
- **Conditional Gradient methods (Frank-Wolfe):** Each iteration linearizes the quadratic term and solves a classical or partial OT subproblem (often via entropic Sinkhorn) to update the coupling. For FPGW, the partial OT subproblem is a linear program with mass constraints [2502.09934], [1811.02834].
- **Sinkhorn-Type Iterative Schemes:** Especially for entropic regularization, iterates alternate between marginal projections and fixed-point updates. Unbalanced OT problems leverage KL penalization and employ unbalanced Sinkhorn [2206.09398], [2411.10204].
- **Majorization-Minimization and Block-Coordinate Descent:** For barycenter and multi-marginal settings, alternating minimization efficiently solves tight bi-convex relaxations [2205.06725].
- **Sliced and Linear Approximations:** The sliced FGW reduces complexity by projecting onto 1D subspaces and leveraging hierarchical lower bounds and Monte Carlo slicing, achieving scalability for large and complex datasets [2508.02364], [2203.04711], [2411.10204].

The table below outlines notable solvers:

| Variant                 | Optimization Method     | Scalability      |
|-------------------------|------------------------|------------------|
| FGW (balanced)          | Frank-Wolfe, Sinkhorn  | $O(n^3)$, up to $O(n^2)$ with entropy |
| FPGW, FUGW              | Partial OT, Unbalanced Sinkhorn | $O(nm\log(nm))$ per iteration |
| Sliced FGW              | Monte-Carlo slicing, 1D OT | $O(L\,n\log n)$ after pre-computation |
| Multi-marginal FGW      | Alternating Sinkhorn   | $O(N\,n^2)$ per iteration |
| Linear FGW (embedding)  | Barycentric projection | $O(N\,m^2K)$ total for $N$ graphs |

## 4. Theoretical Properties: Metrics, Interpolation, and Variance Analysis

FGW is a metric or semi-metric depending on the exponent $q$ and structure:
- Nonnegativity, symmetry always hold.
- Identity of indiscernibles: $d_{FGW}=0$ iff the structured objects are matched by an isometry preserving both feature and structure [2502.09934], [1811.02834], [2306.15963].
- Triangle inequality: exact if $q=1$, else relaxed by factor $2^{q-1}$.
- FGW interpolates between Wasserstein (feature-only, $\alpha=0$) and Gromov-Wasserstein (structure-only, $\alpha=1$).
- Fréchet means/barycenters: existence and closed-form block-coordinate updates enable barycenter computation and clustering [1811.02834], [2202.03813], [2205.06725].

Variance decompositions via Linear Optimal Transport (LOT) provide exact splits into deterministic and residual terms:
\[
FGW_{2,\alpha}^2(\mathcal X,\mathcal Y) = FGW_{2,\alpha}^2(\mathcal X,\mathcal T) + \alpha C_W + (1-\alpha) C_{GW}
\]
where $\mathcal T$ is the barycentric projection. The percent explained by the LOT embedding guides the selection of embedding dimension $n$ for dimension reduction and model building [2411.10204].

## 5. Applications: Graph Learning, Clustering, Alignment, and Data Augmentation

FGW supports a wide class of applications:

- **Graph Classification and Kernel Methods:** FGW distances can be used as kernels (RBF, indefinite, linearized PSD variants) for SVMs and clustering. FGW-based kernels surpass classical graph kernels (WL, SPK, etc.) on benchmarks such as MUTAG, ENZYMES, PROTEINS [1811.02834], [2306.15963], [2203.04711], [2411.10204].
- **FGW Barycenters and Clustering:** FGW’s barycentric averaging produces meaningful prototypical graphs, yielding strong separation in block-model clustering and graph community detection [1811.02834], [2202.03813], [2205.06725].
- **Graph Prediction and Template Learning:** FGW barycenter layers, differentiable via Envelope Theorem, enable end-to-end template learning in graph neural network architectures, achieving competitive or superior empirical results [2205.15733].
- **Video, Time-Series Alignment, and Neuroimaging:** Partial/unbalanced variants robustly handle outlier segments, background/noise, and anatomical discrepancies [2507.15540], [2509.24382], [2206.09398].
- **Graph Mixup Augmentation:** FGWMixup synthesizes graph interpolants in FGW space, improving generalizability of GNNs; scalable solvers (relaxed MD, Sinkhorn) yield fast convergence [2306.15963].
- **Graph Matching, Assignment, and Subgraph Retrieval:** FGW and its extensions (partial, regularized, sliced) provide effective frameworks for assignment problems, subgraph matching in massive graphs, and keypoint correspondence, with high robustness to feature noise [2509.04089], [2406.19767].

## 6. Recent Developments: Edge Features, Multi-Marginal and Sliced FGW

- **Fused Network GW (FNGW):** Incorporates both node and edge features through an additional fused cost on edge-attribute tensors. Proven metric properties, block coordinate solvers, and efficiency gains empirically outperform FGW and other graph kernels when edge attributes matter [2309.16604].
- **Multi-Marginal FGW:** Generalizes to aggregates and interpolates among multiple structured spaces; alternating Sinkhorn and tight bi-convex relaxation enable multi-object barycenter computation [2205.06725].
- **Sliced FGW:** Hierarchical and quadrature-based slicing drastically reduces computational costs, maintains isometric invariance, and achieves pseudo-metric properties suitable for shape retrieval and graph isomorphism testing [2508.02364].

## 7. Practical Guidelines and Empirical Insights

- Trade-off parameter $\alpha$ is critical: in graph tasks, $\alpha\in[0.5,0.95]$ typically balances signal and structure correctly.
- LOT variance decompositions quantify the efficacy of dimensionality reduction and can guide embedding size ($n=10$–$75$ often suffices) [2411.10204].
- Regularization, partial mass matching, and sliced approximations enhance robustness to outliers, noise, and scalability.
- The Python Optimal Transport (POT) library implements principal FGW and related algorithms, facilitating integration and reproducible research [2509.04089], [2205.15733], [2306.15963].

In summary, Fused Gromov-Wasserstein transport supplies a versatile, theoretically rigorous, and empirically validated basis for learning, comparing, and synthesizing structured objects in modern computational science. Advances spanning partial matching, linear and sliced embedding, edge-feature integration, and multi-marginal formulations have established FGW as a foundational tool in graph-based and structured data analysis [2502.09934], [1811.02834], [2205.06725], [2309.16604], [2508.02364], [2411.10204].

Source: https://www.emergentmind.com/topics/fused-gromov-wasserstein-optimal-transport