Structure Graph: Theory & Applications

• Structure graphs are graph-theoretic representations that explicitly encode relationships, hierarchies, and multiscale organization in data.
• They are applied in diverse fields such as robust graph learning, document analysis, crystalline configuration, and neural network architectures.
• Recent methodologies optimize structure graphs via learning refined adjacencies, incorporating high-order connectivity, and employing efficient condensation strategies.

In the literature surveyed here, structure graph does not denote a single standardized object. It denotes a graph-theoretic representation used to make organization explicit when raw observations, flat instances, or a fixed adjacency are insufficient: an induced grammar over repeated graph patterns, a learned adjacency for robust graph learning, a cross-graph matching graph, a document relation graph, a lattice-configuration graph, or a reduced mixed graph preserving conditional independences [0703132, (Jin et al., 2020, Wang et al., 2024, Chen et al., 4 Feb 2025, Yuge, 2016, Wermuth, 2010)]. Across these settings, the common role of a structure graph is to encode relations, proximity, hierarchy, or multiscale organization in a form that can be optimized, analyzed, or composed.

1. Historical emergence and conceptual scope

One early formulation treats structure discovery itself as a graph-compression problem. “Structure induction by lossless graph compression” presents Graphitour, a “novel algorithm” for “structure induction by lossless graph compression,” motivated by vast relational datasets represented as graphs, such as genomic networks. The work states that it extends to graphs grammatical-inference approaches previously applied only to strings, relates bottom-up graph compression to the “maximum cardinality (non-bipartite) maximum cardinality matching problem,” accepts directed graphs and graphs with labeled nodes and arcs, and produces structure that “could be used for representation and classification of graphs” [0703132].

Later graph-learning literature broadens the meaning of structural preservation. A survey of structure-preserving graph transformers defines graph structure preservation in terms of preserving “the connectivity pattern, relative positions, paths, neighborhoods, and higher-order organization” that encode “the underlying relations and proximity” among nodes, and organizes existing strategies into four groups: node feature modulation, context node sampling, graph rewriting, and transformer architecture improvements (Hoang et al., 2024). This suggests that the term is best understood functionally: a structure graph is any graph object introduced to preserve or expose organization that would otherwise be latent or distorted.

2. Learned adjacency as structure graph in robust graph learning

A major contemporary usage treats the structure graph as a learned adjacency matrix that replaces an unreliable observed graph. In Pro-GNN, the observed adjacency $\mathbf A$ is assumed to be poisoned, and robustness is achieved by learning a refined adjacency $\mathbf S \in [0,1]^{N\times N}$ jointly with GNN parameters. The optimization combines fidelity to the observed graph, sparsity, low rank, feature smoothness, and supervised task loss:

$$\| {\bf A} - {\bf S}\|^2_F + \alpha\| {\bf S}\|_1 + \beta \| {\bf S}\|_* + \lambda \, tr( {\bf X}^T\hat{\bf L} {\bf X}) + \gamma \mathcal{L}_{GNN}({\bf \theta}, {\bf S}, {\bf X}, \mathcal{Y}_L),$$

with $\mathbf S=\mathbf S^\top$. The learned structure is therefore a weighted, symmetric, constrained adjacency used directly for message passing rather than a preprocessing artifact (Jin et al., 2020).

A related but methodologically distinct formulation appears in GSEBO, which argues that many GSL methods rely too heavily on local edge-wise heuristics and should instead optimize structure using global mapping information learned by the GNN. It extracts graph structure as a learnable parameter matrix $\mathbf Z$ over observed edges only and rewrites propagation through

$$GSE(\mathbf Z):=\sigma(\mathbf Z)\odot \tilde{\mathbf A}.$$

The model then uses bi-level optimization: the inner level optimizes GNN parameters on training loss, while the outer level optimizes structure on validation loss, so that structure learning is guided by global generalization rather than only local edge evidence (Yin, 2024).

These formulations make a strong conceptual shift. Standard GNN training assumes graph structure is trusted data. Structure-graph learning instead treats graph structure as a latent object to be estimated, reweighted, or denoised.

3. High-order, self-supervised, clustering-oriented, and sparse refinement

Several later works preserve structure not merely by denoising $\mathbf A$, but by incorporating higher-order connectivity or self-supervised signals. “Exploring High-Order Structure for Robust Graph Structure Learning” defines high-order graph structure by powers of the normalized adjacency, $\hat{\mathbf A}^k$, and learns a refined graph with a penalty

$$\sum_{k=1}^{K}\eta_k\|\hat{\mathbf S}-\hat{\mathbf A}^k\|_F^2.$$

The paper states that high-order graph structure denotes the path number between nodes, that it acts as a smoother filter, and that in practice $K=2$ or $\mathbf S \in [0,1]^{N\times N}$0 is used, with $\mathbf S \in [0,1]^{N\times N}$1 in experiments (Yang et al., 2022).

SUBLIME moves to fully unsupervised graph structure learning. It distinguishes a learner view $\mathbf S \in [0,1]^{N\times N}$2 and an anchor view $\mathbf S \in [0,1]^{N\times N}$3, trains them by node-level contrastive learning, and updates the anchor by

$\mathbf S \in [0,1]^{N\times N}$4

This framework explicitly targets “unsupervised graph structure learning,” where the optimized graph topology is learned from the data itself rather than from labels (Liu et al., 2022).

GSR separates structure estimation from downstream supervision by a pretrain–refine–finetune pipeline. A multi-view contrastive pretraining phase performs intra- and inter-view link prediction; the graph is then refined by adding top $\mathbf S \in [0,1]^{N\times N}$5 non-edges with highest probabilities and removing $\mathbf S \in [0,1]^{N\times N}$6 existing edges with least probabilities; the refined graph remains static during finetuning, which is the source of the method’s efficiency and scalability gains (Zhao et al., 2022).

In unsupervised graph clustering, HoLe defines structure learning as homophily enhancement. Its two central edit operators are intra-cluster Edge Recovery and inter-cluster Edge Removal, driven by Hierarchical Correlation Estimation and Cluster-aware Sparsification. The graph update

$\mathbf S \in [0,1]^{N\times N}$7

is explicitly designed to increase edge homophily and improve clustering without labels (Gu et al., 2023).

A more restrictive interpretation of informative structure appears in the Graph Lottery Ticket Hypothesis, which states that any graph contains a sparse subset of edges such that graph learning algorithms can match the full graph’s performance on that subgraph. The proposed kTree method constructs such tickets by taking unions of random spanning trees, and the paper reports that performance can often be “matched or even exceeded on graphs with the average degree as low as 5” (Tsitsulin et al., 2023). Here the structure graph is a sparse backbone rather than a dense refined adjacency.

4. Cross-graph, neural-architectural, and document relation structures

In another usage, the structure graph is not a denoised adjacency but an explicit graphization of computation or correspondence. “Graph Structure of Neural Networks” defines a relational graph $\mathbf S \in [0,1]^{N\times N}$8 whose nodes carry features and whose edges specify which nodes exchange information. One round of message exchange is

$\mathbf S \in [0,1]^{N\times N}$9

and layers correspond to rounds of message exchange on this graph (You et al., 2020). In this setting, the structure graph is the neural architecture itself.

In graph similarity learning, SEGMN introduces the assignment graph $$\| {\bf A} - {\bf S}\|^2_F + \alpha\| {\bf S}\|_1 + \beta \| {\bf S}\|_* + \lambda \, tr( {\bf X}^T\hat{\bf L} {\bf X}) + \gamma \mathcal{L}_{GNN}({\bf \theta}, {\bf S}, {\bf X}, \mathcal{Y}_L),$$0, which is explicitly a graph over candidate cross-graph node pairs. Nodes of $$\| {\bf A} - {\bf S}\|^2_F + \alpha\| {\bf S}\|_1 + \beta \| {\bf S}\|_* + \lambda \, tr( {\bf X}^T\hat{\bf L} {\bf X}) + \gamma \mathcal{L}_{GNN}({\bf \theta}, {\bf S}, {\bf X}, \mathcal{Y}_L),$$1 are pairs $$\| {\bf A} - {\bf S}\|^2_F + \alpha\| {\bf S}\|_1 + \beta \| {\bf S}\|_* + \lambda \, tr( {\bf X}^T\hat{\bf L} {\bf X}) + \gamma \mathcal{L}_{GNN}({\bf \theta}, {\bf S}, {\bf X}, \mathcal{Y}_L),$$2, and two such nodes are adjacent iff $$\| {\bf A} - {\bf S}\|^2_F + \alpha\| {\bf S}\|_1 + \beta \| {\bf S}\|_* + \lambda \, tr( {\bf X}^T\hat{\bf L} {\bf X}) + \gamma \mathcal{L}_{GNN}({\bf \theta}, {\bf S}, {\bf X}, \mathcal{Y}_L),$$3 and $$\| {\bf A} - {\bf S}\|^2_F + \alpha\| {\bf S}\|_1 + \beta \| {\bf S}\|_* + \lambda \, tr( {\bf X}^T\hat{\bf L} {\bf X}) + \gamma \mathcal{L}_{GNN}({\bf \theta}, {\bf S}, {\bf X}, \mathcal{Y}_L),$$4. Graph convolution on this assignment graph propagates confidence among structurally compatible matches and refines node-pair similarity scores (Wang et al., 2024). The structure graph here is second-order: it is a graph over the matching space rather than over either input graph alone.

Document understanding provides yet another variant. In graph-based Document Structure Analysis, the target is a document structure graph

$$\| {\bf A} - {\bf S}\|^2_F + \alpha\| {\bf S}\|_1 + \beta \| {\bf S}\|_* + \lambda \, tr( {\bf X}^T\hat{\bf L} {\bf X}) + \gamma \mathcal{L}_{GNN}({\bf \theta}, {\bf S}, {\bf X}, \mathcal{Y}_L),$$5

whose nodes are instance-level layout elements and whose edges are multi-relational spatial and logical links. The GraphDoc dataset contains 80K document images and 4.13M relation annotations across 11 layout categories and eight relation categories—Up, Down, Left, Right, Parent, Child, Sequence, and Reference—and the DRGG benchmark reaches 57.6% at $$\| {\bf A} - {\bf S}\|^2_F + \alpha\| {\bf S}\|_1 + \beta \| {\bf S}\|_* + \lambda \, tr( {\bf X}^T\hat{\bf L} {\bf X}) + \gamma \mathcal{L}_{GNN}({\bf \theta}, {\bf S}, {\bf X}, \mathcal{Y}_L),$$6 (Chen et al., 4 Feb 2025). Here the structure graph converts a flat page layout into a directed, typed, heterogeneous, multi-relational graph.

5. Structural encodings in physical systems and condensed representations

In materials science, a structure graph is used to encode crystalline configuration. “Graph Representation for Configurational Properties of Crystalline Solids” defines, for a chosen pair shell $$\| {\bf A} - {\bf S}\|^2_F + \alpha\| {\bf S}\|_1 + \beta \| {\bf S}\|_* + \lambda \, tr( {\bf X}^T\hat{\bf L} {\bf X}) + \gamma \mathcal{L}_{GNN}({\bf \theta}, {\bf S}, {\bf X}, \mathcal{Y}_L),$$7, a graph

$$\| {\bf A} - {\bf S}\|^2_F + \alpha\| {\bf S}\|_1 + \beta \| {\bf S}\|_* + \lambda \, tr( {\bf X}^T\hat{\bf L} {\bf X}) + \gamma \mathcal{L}_{GNN}({\bf \theta}, {\bf S}, {\bf X}, \mathcal{Y}_L),$$8

where vertices are lattice sites and edges are symmetry-equivalent neighboring pairs. The associated matrices $$\| {\bf A} - {\bf S}\|^2_F + \alpha\| {\bf S}\|_1 + \beta \| {\bf S}\|_* + \lambda \, tr( {\bf X}^T\hat{\bf L} {\bf X}) + \gamma \mathcal{L}_{GNN}({\bf \theta}, {\bf S}, {\bf X}, \mathcal{Y}_L),$$9, $\mathbf S=\mathbf S^\top$0, and $\mathbf S=\mathbf S^\top$1 connect graph spectra to generalized Ising model correlations; for pair correlations,

$\mathbf S=\mathbf S^\top$2

The paper’s main claim is that the spectrum of these graphs and of linear combinations such as $\mathbf S=\mathbf S^\top$3 exposes higher-order closed links that are not visible in low-dimensional pair-correlation descriptions alone (Yuge, 2016).

In graph condensation, the structure graph becomes a synthetic adjacency built for efficient training. GCSR reconstructs a condensed graph by a self-expressive model regularized by both an original-graph prior and historical structure:

$\mathbf S=\mathbf S^\top$4

$\mathbf S=\mathbf S^\top$5

The resulting synthetic graph is intended to be interpretable because it reflects self-representation among condensed nodes together with class-wise structural information inherited from the original graph (Liu et al., 2024).

A related representational idea appears in SS-GDE, where each input graph is embedded relative to a learned base graph dictionary that is adapted to the input by Bernoulli node selection, producing an adapted graph dictionary, and graph–dictionary correspondence is then measured by multi-sensitivity Wasserstein encoding (Liu et al., 2023). In this usage, the structure graph is neither a single adjacency nor a sparse subgraph, but a set of adaptive graph keys against which structural correspondence is measured.

6. Summary graphs, product structure, and hierarchical lineages

In probabilistic graphical modeling, summary graphs are mixed graphs for reduced distributions obtained after marginalizing and conditioning. A summary graph $\mathbf S=\mathbf S^\top$6 has node set $\mathbf S=\mathbf S^\top$7 partitioned into $\mathbf S=\mathbf S^\top$8, with a mixture of a directed acyclic graph and a covariance graph within $\mathbf S=\mathbf S^\top$9, a concentration graph within $\mathbf Z$0, and arrows from $\mathbf Z$1 to $\mathbf Z$2. The paper emphasizes that summary graphs preserve the independence structure implied by the generating process and remain closed under further marginalizing and conditioning (Wermuth, 2010). The structure graph here is a reduced graphical object that records which dependences and independences survive reduction.

In structural graph theory, graph structure can itself be described by product decompositions. “Notes on Graph Product Structure Theory” surveys the theorem that every planar graph is a subgraph of

$\mathbf Z$3

for some graph $\mathbf Z$4 of treewidth at most 8 and some path $\mathbf Z$5, together with generalizations to graphs on surfaces, minor-closed classes, various non-minor-closed classes, and graph classes with polynomial growth (Dvořák et al., 2020). In this setting, structure is expressed by factorization into simpler graph components rather than by learning or refinement.

A still more explicitly hierarchical formulation appears in graph lineages. “Graph Lineages and Skeletal Graph Products” defines structured graph lineages ordered by level number, with exponentially increasing numbers of vertices and edges, bipartite graphs connecting successive levels, prolongation maps and process-derived distances between graphs at successive levels, and low-cost skeletal variants of graph products and type constructors on graded graphs. The paper further introduces thickening and escalation as unary operators producing multiscale graph lineages useful for hierarchical model architectures and multigrid methods (Mjolsness et al., 31 Jul 2025). Here the structure graph is fundamentally multilevel: not one graph, but an organized sequence of graphs with explicit cross-level relations.

Taken together, these usages show that structure graph is best understood as a family of graph-theoretic constructions for making organization explicit. In some contexts it is a refined adjacency optimized for robustness or clustering; in others it is a graph over neural communication, candidate correspondences, document elements, lattice neighborhoods, reduced probabilistic dependences, product factors, or successive scales. The persistent theme is that structure is treated not as incidental decoration on data, but as an object that can be induced, preserved, compressed, transported, sparsified, or composed.