Graph Diffusion Network

• Graph-Diffusion Network is a graph neural architecture that replaces one-hop message passing with explicit diffusion processes to capture richer neighborhood information.
• It integrates multi-hop discrete aggregation and continuous-time PDE formulations to mitigate over-smoothing and enhance representation learning.
• The approach supports diverse formulations—including equilibrium, invertibility, and generative diffusion—enabling robust applications in node classification, link prediction, and topology generation.

A graph-diffusion network is a graph neural network whose propagation mechanism is explicitly formulated as a diffusion process on a graph, rather than only as finite one-hop message passing. In this literature, diffusion can appear as a generalized random-walk or Laplacian operator, a closed-form equilibrium smoother, a continuous-time ODE or PDE, an implicit fixed-point layer, an invertible residual process for source recovery, or a denoising diffusion probabilistic model over graph-structured variables (Gasteiger et al., 2019, Jiang et al., 2018, Cui et al., 2024).

1. Diffusion as the organizing principle

A common formalization starts from a generalized graph diffusion operator

$S=\sum_{k=0}^{\infty}\theta_k T^k,$

where $T$ is a transition matrix derived from the graph and $\theta_k$ are non-negative coefficients. In "Diffusion Improves Graph Learning" this construction subsumes Personalized PageRank and the heat kernel, and is explicitly related to polynomial spectral filters; the resulting operator acts as a low-pass filter while remaining usable in spatial GNN pipelines through sparsified diffused adjacencies (Gasteiger et al., 2019). In that sense, a graph-diffusion network is a network whose effective neighborhood system is defined by diffusion rather than by the raw one-hop adjacency.

A second canonical viewpoint appears in "Graph Diffusion-Embedding Networks", where each layer first computes an equilibrium diffusion of features and then applies a learnable embedding map,

$$X^{(k)}=\sigma\big(\mathcal{H}_d(A,X^{(k-1)})W^{(k)}\big).$$

The diffusion operator $\mathcal{H}_d$ can be Laplacian diffusion, random walk with restart, or normalized Laplacian diffusion, all given in closed form. This makes the diffusion step itself the solution of a regularized optimization problem, rather than a heuristic local aggregation rule (Jiang et al., 2018).

A third line of work treats standard GCN propagation as a power iteration. "Graph Neural Diffusion Networks for Semi-supervised Learning" rewrites simplified deep propagation as

$H^{(k)}=\widetilde{W}^kX\Theta,$

and uses this to distinguish under-smoothing from over-smoothing: small $k$ does not propagate sparse label information far enough, while large $k$ drives representations toward the dominant eigenvector of $\widetilde{W}$ (Ye et al., 2022). This perspective motivates architectures that control diffusion depth directly instead of only increasing network depth.

2. Discrete multi-hop architectures

One important family keeps diffusion discrete and layerwise, but replaces single-hop propagation with explicit multi-hop aggregation. "Adaptive Graph Diffusion Networks" defines a diffusion layer through smaller multi-hop node representations $\overline{A}^kH$ combined with learnable generalized weighting coefficients. Its two adaptive mechanisms, Hop-wise Attention and Hop-wise Convolution, make hop aggregation node-wise or channel-wise while maintaining moderate complexity and runtime; the method was evaluated on OGB node classification and link prediction benchmarks and reported top-1 results on ogbn-arxiv, ogbn-proteins, and ogbl-ddi at submission time (Sun et al., 2020).

"GND-Nets" pursues a related objective for sparsely labeled semi-supervised learning. Rather than stacking many GCN layers, it computes

$T$0

inside a shallow architecture and learns how to aggregate these terms with a Single-Layer Perceptron. The resulting neural diffusion

$T$1

is presented as a learnable generalization of classical linear and nonlinear graph diffusions, with the explicit goal of mitigating both under-smoothing and over-smoothing (Ye et al., 2022).

"GIDN: A Lightweight Graph Inception Diffusion Network for High-efficient Link Prediction" extends this multi-hop idea to different feature spaces. Its inception module applies several diffusion branches in parallel, each with its own projection and hop weighting, so that multi-scale diffusion is captured by width rather than depth. The paper reports Hits@50 values of $T$2 for AGDN, $T$3 for PLNLP, and $T$4 for GIDN on ogbl-collab, and describes the gain over AGDN as about 11% (Wang et al., 2022).

A recurrent misconception is that graph diffusion is necessarily a fixed preprocessing step. The literature does contain preprocessing-based systems such as GDC, but it also contains per-layer diffusion operators with learnable hop weighting, branch-specific feature spaces, and residual transforms (Gasteiger et al., 2019, Sun et al., 2020).

3. Continuous-time, PDE, and variational formulations

Another major branch formulates graph-diffusion networks as continuous-time dynamics. In "Graph neural Aggregation-diffusion with Metastability", node states evolve according to an aggregation-diffusion ODE,

$T$5

which combines nonlinear diffusion with nonlocal aggregation induced by interaction potentials. The paper argues that such dynamics produce metastable clustered states rather than rapid collapse to a single equilibrium, and proves a mitigation result for over-smoothing under a logarithmic interaction kernel (Cui et al., 2024).

"Graph Neural Convection-Diffusion with Heterophily" modifies the diffusion equation by adding a convection term,

$T$6

where the velocity field is parameterized from feature differences,

$T$7

Here diffusion models homophily-compatible smoothing, while convection models directed transport associated with heterophily (Zhao et al., 2023).

A variational counterpart appears in "Re-Think and Re-Design Graph Neural Networks in Spaces of Continuous Graph Diffusion Functionals". That work identifies standard layer-by-layer embedding models with an $T$8-norm integral functional of graph gradients, relates this to over-smoothing, and replaces it with a total-variation functional plus a selective mechanism implemented through diffusion-clip layers. It also introduces a neural transport equation and an adversarial flow-prediction framework on graphs (Dan et al., 2023).

These formulations show that graph diffusion need not mean passive heat flow. This suggests that the term covers a broader class of graph dynamics in which smoothing, transport, aggregation, and regularization are all encoded as structured operators on node states.

4. Equilibrium, invertibility, and constraints

Graph-diffusion networks also appear in inverse and equilibrium settings. "Implicit Graph Neural Diffusion Networks: Convergence, Generalization, and Over-Smoothing" defines an implicit layer by the fixed-point equation

$T$9

where $\theta_k$0 is either a random-walk Laplacian or a learned parameterized graph Laplacian. The paper analyzes over-smoothing during training and inference, proves uniqueness and geometric convergence when $\theta_k$1 exceeds the largest eigenvalue of the Laplacian, and gives transductive generalization bounds tied to the diffusion spectrum (Fu et al., 2023).

"An Invertible Graph Diffusion Neural Network for Source Localization" pushes the idea further by making a graph diffusion model explicitly invertible. Its residual construction

$\theta_k$2

admits fixed-point inversion when $\theta_k$3 and $\theta_k$4 are contractive. The framework adds an error compensation module and validity-aware layers that enforce source constraints through unrolled optimization, with a convergence result for linear constraints (Wang et al., 2022). This directly contradicts the common assumption that diffusion models are only forward simulators of spreading processes.

"An Active Diffusion Neural Network for Graphs" addresses equilibrium from a different angle. It introduces source terms into the diffusion process,

$\theta_k$5

and derives the closed-form infinite-diffusion solution

$\theta_k$6

The source terms are intended to preserve node distinctiveness while still capturing global structure (Jiang, 22 Oct 2025).

5. Beyond Euclidean smoothing: manifolds, self-supervision, and generative diffusion

The diffusion idea also extends beyond Euclidean node features. "Manifold GCN" defines a manifold-valued graph Laplacian by

$\theta_k$7

and discretizes the corresponding diffusion equation through tangent-space Euler steps and exponential-map updates. The resulting diffusion layer and tangent MLP are equivariant under both node permutations and feature-manifold isometries, and the paper proves global existence under a convexity condition on the data’s geodesic ball and $\theta_k$8 (Hanik et al., 2024).

In self-supervised graph learning, "Self-Contrastive Graph Diffusion Network" couples an Attentional Module with a Diffusion Module governed by

$\theta_k$9

and trains the resulting embeddings with a Laplacian-based self-contrastive loss and redundancy reduction. Its defining claim is that it is augmentation-free, does not require pre-training, and derives positive and negative pairs directly from graph structure and feature-based $$X^{(k)}=\sigma\big(\mathcal{H}_d(A,X^{(k-1)})W^{(k)}\big).$$0NN information (Ma et al., 2023).

A very different use of the term appears in generative modeling. "NetDiff" is a graph denoising diffusion probabilistic model for ad hoc wireless topology generation, where the reverse model is a graph transformer enhanced with Cross-Attentive Modulation tokens and partial diffusion for topology evolution over time (Marcoccia et al., 2024). "Graph Diffusion Network for Drug-Gene Prediction" uses a parallel DDPM-like diffusion model on gene embeddings to generate hard negatives conditioned on drug embeddings, while the graph encoder itself relies on meta-path-based homogeneous graph learning and bipartite aggregation (Wu et al., 13 Feb 2025). These systems indicate that graph diffusion can denote either propagation on a graph or denoising diffusion in graph-conditioned latent space. A plausible implication is that the phrase "graph-diffusion network" now names a family of design patterns rather than a single operator class.

6. Applications, recurring problems, and interpretive boundaries

Applications span semi-supervised node classification, link prediction, clustering, source localization, graph classification, topology generation, and biomedical association prediction. The same label therefore covers APPNP-like preprocessing, equilibrium feature smoothers, inception-style multi-hop encoders, invertible source-recovery models, manifold-valued diffusion layers, and DDPM-based graph generators (Gasteiger et al., 2019, Wang et al., 2022, Wang et al., 2022).

Across these variants, several tensions recur. The first is over-smoothing: passive or repeated diffusion tends to drive representations toward equilibrium, which is explicitly analyzed in GND-Nets, GRADE, DIGNN, and ADGNN (Ye et al., 2022, Cui et al., 2024, Fu et al., 2023). The second is efficiency: GDC emphasizes sparsified preprocessing without changing downstream asymptotic complexity; AGDN stresses moderate complexity and runtime; IVGD uses linearized augmented Lagrangian layers for scalable constrained inference; NetDiff adds partial diffusion to avoid full regeneration when graphs evolve (Gasteiger et al., 2019, Sun et al., 2020, Wang et al., 2022, Marcoccia et al., 2024). The third is task specificity: some models are built for homophilic node classification, some for heterophily, some for inverse diffusion, and some for generative design. This suggests that no single diffusion formalism is uniformly dominant across graph tasks.

A final interpretive boundary concerns what counts as “diffusion.” In some works it is a weighted sum of powers of a transition matrix; in others it is the equilibrium of a regularized functional; elsewhere it is a continuous-time aggregation-diffusion PDE, an invertible residual map, or a denoising diffusion probabilistic process over embeddings or edges. The technically stable core is narrower: graph-diffusion networks make the transport of information over graph structure an explicit object of model design, analysis, or inversion, rather than an incidental consequence of stacked one-hop layers.