DualGNN: Dual-Formulation GNN Architectures

• DualGNN is defined as a family of architectures featuring two complementary operators (e.g., graphs, filters, or channels) to address issues like oversmoothing and incomplete data.
• These models employ dual-view propagation, dual-operator filtering, or dual-branch specialization to optimize complementary objectives in tasks such as node classification and imaging.
• Empirical studies report significant performance gains using dualGNN approaches, although challenges remain in computational overhead and consistent terminology.

The label dualGNN is not used uniformly across the graph-learning literature. The exact name appears in at least two distinct contexts: a semi-supervised node-classification framework that couples a primary GNN with an auxiliary GNN on a spectrally constructed graph (Alchihabi et al., 2021), and an autoregressive message-passing GNN for sampling fine, regular triangulations of convex polytopes (MacFadden, 26 May 2026). Closely related usage also appears in works that adopt different formal names while instantiating the same broad design idea: two coupled graph-based operators, views, spaces, or branches trained within a single architecture, as in FB-GNNs, DualHGNN, MDS-GNN, DuoGNN, DG-NLOS, and DualEquiNet (Luan et al., 2022, Liao et al., 2023, Yuan et al., 2024, Mancini et al., 2024, Su et al., 27 Feb 2025, Xu et al., 10 Jun 2025). In current practice, the term is therefore best understood as a family of dual-formulation GNN architectures rather than a single canonical model.

1. Terminology and scope

A persistent source of ambiguity is that “dual” denotes different pairings in different papers. In Dual GNNs, the duality is between a primary GNN operating on the original graph and an auxiliary GNN operating on a graph constructed from fine-grained spectral clustering (Alchihabi et al., 2021). In FB-GNNs, the duality is between low-pass aggregation and high-pass diversification channels (Luan et al., 2022). In DualHGNN, it is between hypergraph structure learning and hypergraph representation learning, and the paper explicitly states that it consistently uses the name DualHGNN rather than “DualGNN” (Liao et al., 2023). In MDS-GNN, the duality is a mutual dual-stream design in which structure reconstructs features and reconstructed features rebuild structure (Yuan et al., 2024). In DuoGNN, the two branches decouple homophilic and heterophilic interactions (Mancini et al., 2024). In DG-NLOS, the branches are albedo-focused and depth-focused (Su et al., 27 Feb 2025). In DualEquiNet, the two spaces are Euclidean and Spherical Harmonics (Xu et al., 10 Jun 2025).

This variability matters because the dualGNN idea is architectural rather than task-specific. The dual components may be two graphs, two filters, two propagation spaces, two message types, or two coupled optimization variables. A plausible implication is that “dual” in this literature functions less as a taxonomic term and more as a design principle for pairing complementary inductive biases.

Model	Duality	Task
“Dual GNNs” (Alchihabi et al., 2021)	original graph + auxiliary spectral-clustering graph	semi-supervised node classification
FB-GNNs (Luan et al., 2022)	low-pass aggregation + high-pass diversification	node classification
“DualHGNN” (Liao et al., 2023)	hypergraph structure learning + density-aware representation learning	semi-supervised node classification
“MDS-GNN” (Yuan et al., 2024)	structure-to-feature + feature-to-structure streams	incomplete graphs
“DuoGNN” (Mancini et al., 2024)	homophilic branch + heterophilic branch	node classification
“DG-NLOS” (Su et al., 27 Feb 2025)	albedo-focused branch + depth-focused branch	confocal NLOS imaging

2. Recurring architectural patterns

Across these models, several structural motifs recur. The first is dual-view propagation. In the 2021 Dual GNN framework, the primary encoder $f_\Theta(X,A)$ learns on the observed graph, while an auxiliary branch builds a new adjacency $A_{sc}$ from soft spectral cluster assignments and applies a second GNN on that graph (Alchihabi et al., 2021). In MDS-GNN, the first stream uses incomplete structure $A'$ to reconstruct missing features, after which reconstructed features are used to build an augmented global graph $\hat A$ for APPNP-style propagation (Yuan et al., 2024). D$^2$PT follows a related pattern at the level of propagation channels, diffusing once on the incomplete input graph and once on a constructed global semantic graph, then aligning class prototypes across the two channels (Liu et al., 2023).

The second motif is dual-operator filtering. FB-GNNs formalize this most explicitly by pairing a low-pass channel with its complementary high-pass channel. The low-pass branch captures smooth graph signals, while the high-pass branch preserves differences between a node and its neighbors, counteracting oversmoothing (Luan et al., 2022). The same general logic appears in DuoGNN, where one branch is specialized for homophilic aggregation and the other for heterophilic long-range interactions on a condensed graph (Mancini et al., 2024).

A third motif is dual-branch specialization by semantic role. DG-NLOS separates albedo reconstruction from depth reconstruction, trains the albedo-focused branch first, freezes it, and then trains a depth-focused branch using a depth-focused masking rule (Su et al., 27 Feb 2025). DualEquiNet similarly separates Euclidean local geometry from Spherical Harmonics structure, then couples them by bidirectional cross-space message passing and Cross-Space Interaction Pooling (Xu et al., 10 Jun 2025). In both cases, the dual design is intended to reduce interference between signals that are useful but not identical.

A fourth motif is joint optimization of complementary objectives. DualHGNN combines hypergraph learning and supervised classification in a single loss, while its density-aware attention mechanism modulates node–hyperedge–node propagation on the learned hypergraph (Liao et al., 2023). In weather post-processing, dualGNN refers to a GraphSAGE model trained with a composite loss that combines the Energy Score and the Variogram Score, so that both marginal calibration and multivariate dependence are optimized end-to-end (Lakatos, 2 Sep 2025). This suggests that “dual” can refer not only to architecture but also to the training criterion.

3. Representative mathematical formulations

Several representative equations clarify how different dualGNN variants operationalize the dual idea.

In Dual GNNs, the core training objective is

$$\mathcal{L} = \mathcal{L}_{sup}^{(p)} + \mathcal{L}_{sup}^{(a)} + \mathcal{L}_{sc},$$

where the primary supervised loss, auxiliary supervised loss, and relaxed spectral clustering loss are optimized jointly. The auxiliary adjacency is obtained by thresholding Pearson correlations between soft cluster assignment vectors, which makes the auxiliary graph denser and more globally informed than the observed graph (Alchihabi et al., 2021).

In FB-GNNs, the canonical two-channel layer is

$$H^{(\ell)}_{\text{low}}=\hat{A}\,\mathrm{ReLU}(H^{(\ell-1)}W^{(\ell-1)}_{\text{low}}),\quad H^{(\ell)}_{\text{high}}=(I-\hat{A})\,\mathrm{ReLU}(H^{(\ell-1)}W^{(\ell-1)}_{\text{high}}),$$

followed by a learned fusion of the two channels. The relation $\hat A + \hat L = I$ gives a complementary low-pass/high-pass filterbank interpretation (Luan et al., 2022).

In DualHGNN, the hypergraph-learning and classification terms are combined as

$$\mathcal{L}=\mathcal{L}_{HGL} + \lambda \mathcal{L}_{CE}.$$

The learned incidence matrix $H$ is averaged across views and merged with the initial hypergraph $A_{sc}$0 as

$A_{sc}$1

after which density-aware hypergraph attention performs node–hyperedge–node propagation (Liao et al., 2023).

In MDS-GNN, the total loss is

$A_{sc}$2

combining the classification loss on the augmented stream, the classification loss on the original stream, masked reconstruction of observed features, and node-level contrastive alignment between the two streams (Yuan et al., 2024).

In DG-NLOS, the duality is enforced by stage-wise optimization rather than a single simultaneous loss. Stage 1 minimizes $A_{sc}$3; stage 2 freezes the albedo branch and minimizes $A_{sc}$4 (Su et al., 27 Feb 2025). In weather dualGNN, the corresponding dual loss is

$A_{sc}$5

with $A_{sc}$6, so the model directly learns ensemble samples with both global multivariate fidelity and dependence sensitivity (Lakatos, 2 Sep 2025).

A plausible implication is that dualGNN architectures are unified less by any one propagation rule than by a common optimization pattern: two complementary objects are learned together, with explicit coupling terms preventing the branches from collapsing into redundant copies.

4. Domain-specific instantiations

The dualGNN idea first appeared most directly in semi-supervised node classification under limited supervision. The 2021 Dual GNNs paper targeted the regime of scarce labels and noisy graph connectivity, using spectral clustering to construct an auxiliary graph that compensates for missing or corrupted structure (Alchihabi et al., 2021). Closely related node-classification work includes FB-GNNs for mixed-frequency graph signals (Luan et al., 2022), DualHGNN for multi-view hypergraph structure learning and density-aware attention (Liao et al., 2023), D$A_{sc}$7PT for graph learning with weak information (Liu et al., 2023), MDS-GNN for simultaneous feature and structure incompleteness (Yuan et al., 2024), DuoGNN for homophily–heterophily interaction decoupling (Mancini et al., 2024), and DPGNN for node-specific multi-step propagation in topology and feature spaces (Zhou et al., 2021).

In imaging and scientific inverse problems, the same principle acquires a task-specific interpretation. DG-NLOS uses a dual-branch graph feature learner for confocal NLOS imaging, where dense 3D grid features are converted to sparse graph structures and reconstructed in separate albedo and depth branches (Su et al., 27 Feb 2025). DM-GNN for visual question answering, as described in the supplied material, uses an object-centric graph and a relation-centric graph, together with enhanced message passing between objects, relations, and attributes (Li et al., 2022).

In geometric learning, DualEquiNet constructs complementary Euclidean and Spherical Harmonics representations for large biomolecules, coupling the two with bidirectional cross-space message passing and a hierarchical pooling mechanism (Xu et al., 10 Jun 2025). DuSEGO instead applies dual second-order equivariant graph ODEs to both node embeddings and node coordinates, so “dual” refers to the simultaneous continuous-time evolution of feature and geometric states (Wang et al., 2024). In constrained optimization, dualGNN has also been used for two coupled unrolled GNNs that mirror dual ascent, with one network updating primal variables and the other updating dual multipliers layer by layer (Hadou et al., 21 Sep 2025).

The exact lowercase name dualGNN appears in 2026 in a distinct combinatorial setting: an autoregressive message-passing GNN for sampling fine, regular triangulations of convex lattice polytopes. There the model operates on a generalized dual graph whose edges are labeled by signed circuits from oriented matroid theory; the labels are described as both necessary and sufficient for exposing regularity (MacFadden, 26 May 2026). This is the most literal current use of the name, but it is not the dominant historical meaning.

5. Empirical behavior and reported gains

The empirical record for dualGNN-style models is strongest in node classification under graph defects and in settings where a single propagation channel is known to be inadequate. In the original Dual GNNs paper, Dual GCN improved from $A_{sc}$8 to $A_{sc}$9 on Cora and from $A'$0 to $A'$1 on CiteSeer with $A'$2 labels per class, and under $A'$3 edge deletion improved from $A'$4 to $A'$5 on Cora and from $A'$6 to $A'$7 on CiteSeer (Alchihabi et al., 2021). In FB-GNNs, the reported gains were especially large on heterophily-prone datasets: FB-GAT improved over GAT by $A'$8 on Cornell, $A'$9 on Wisconsin, and $\hat A$0 on Texas (Luan et al., 2022).

DualHGNN reported $\hat A$1 on Scene15 with $\hat A$2 labels, $\hat A$3 on CIFAR-10 with $\hat A$4 labels, and $\hat A$5 on MNIST with $\hat A$6 labels, with ablations showing that removing hypergraph learning or density-aware attention reduced performance (Liao et al., 2023). MDS-GNN achieved $\hat A$7 on Cora, $\hat A$8 on Texas, $\hat A$9 on Chameleon, and $^2$0 on Amazon Photo under the paper’s incomplete-feature and incomplete-structure setting, outperforming GCN, GAT, PCFI, SimP-GCN, and often D2PT (Yuan et al., 2024). DuoGNN reported $^2$1 accuracy on MedMNIST Organ-S and $^2$2 on Organ-C, compared with $^2$3 and $^2$4 for GCN (Mancini et al., 2024).

Outside node classification, the same pattern persists. DG-NLOS reported PSNR $^2$5 dB, SSIM $^2$6, RMSE $^2$7, runtime $^2$8 s, and GPU memory $^2$9 GB on $$\mathcal{L} = \mathcal{L}_{sup}^{(p)} + \mathcal{L}_{sup}^{(a)} + \mathcal{L}_{sc},$$0 data, outperforming classical inversions and recent deep methods while remaining lighter than several strong baselines (Su et al., 27 Feb 2025). In multivariate weather post-processing, the ES+VS dualGNN consistently outperformed all empirical copula-based post-processed forecasts on the two datasets considered, and for WRF irradiance its learned rank-order structure yielded a more effective restoration of spatial relationships than either the raw NWP ensemble or historical observational rank structures (Lakatos, 2 Sep 2025). In combinatorial geometry, the 2026 dualGNN is described as the most uniform FRT sampler tested on unseen polygons with $$\mathcal{L} = \mathcal{L}_{sup}^{(p)} + \mathcal{L}_{sup}^{(a)} + \mathcal{L}_{sc},$$1, with a model of approximately $$\mathcal{L} = \mathcal{L}_{sup}^{(p)} + \mathcal{L}_{sup}^{(a)} + \mathcal{L}_{sc},$$2k parameters that trains in approximately $$\mathcal{L} = \mathcal{L}_{sup}^{(p)} + \mathcal{L}_{sup}^{(a)} + \mathcal{L}_{sc},$$3 hours on a single consumer GPU (MacFadden, 26 May 2026).

These results do not establish a universal superiority of any single dual architecture. They do, however, show a repeated empirical pattern: when the task contains two complementary but non-redundant sources of structure, explicitly modeling both often improves calibration, stability, or expressive power.

6. Limitations, misconceptions, and outlook

The first misconception is terminological: dualGNN is not a standardized model name. The literature contains an exact “Dual GNNs” framework (Alchihabi et al., 2021), an exact “dualGNN” triangulation sampler (MacFadden, 26 May 2026), and many non-identical models with different official names that are nevertheless dual in design. The most notable clarification is DualHGNN, whose paper explicitly states that it does not introduce or use the term “DualGNN” and consistently uses the name DualHGNN (Liao et al., 2023).

The second misconception is that “dual” always means two graphs. In fact, it can mean two filters (Luan et al., 2022), two streams (Yuan et al., 2024), two semantic branches (Su et al., 27 Feb 2025), two spaces (Xu et al., 10 Jun 2025), two propagation channels (Liu et al., 2023), or two coupled optimization variables (Hadou et al., 21 Sep 2025). A plausible implication is that future uses of the term will remain heterogeneous unless a narrower convention emerges.

A recurrent limitation is computational overhead. In Dual GNNs, building $$\mathcal{L} = \mathcal{L}_{sup}^{(p)} + \mathcal{L}_{sup}^{(a)} + \mathcal{L}_{sc},$$4 from Pearson correlations is $$\mathcal{L} = \mathcal{L}_{sup}^{(p)} + \mathcal{L}_{sup}^{(a)} + \mathcal{L}_{sc},$$5 and becomes the bottleneck on large graphs (Alchihabi et al., 2021). In MDS-GNN, exact kNN construction and the full-batch contrastive denominators are $$\mathcal{L} = \mathcal{L}_{sup}^{(p)} + \mathcal{L}_{sup}^{(a)} + \mathcal{L}_{sc},$$6 (Yuan et al., 2024). DualHGNN adds multi-view similarity computation, incidence-matrix storage, and density-aware attention overhead (Liao et al., 2023). DG-NLOS requires tuning dynamic $$\mathcal{L} = \mathcal{L}_{sup}^{(p)} + \mathcal{L}_{sup}^{(a)} + \mathcal{L}_{sc},$$7, neighbor selection, and dilation choices (Su et al., 27 Feb 2025). The 2026 dualGNN offers a guarantee that every rollout produces a fine triangulation only in $$\mathcal{L} = \mathcal{L}_{sup}^{(p)} + \mathcal{L}_{sup}^{(a)} + \mathcal{L}_{sc},$$8D, and regularity is learned rather than mathematically enforced (MacFadden, 26 May 2026).

The broader outlook is therefore twofold. On one hand, dualGNN-style designs have become a recurring answer to oversmoothing, incomplete data, weak supervision, dependence modeling, and multi-scale geometry. On the other hand, the term itself remains semantically overloaded. For research use, precision requires specifying the actual paper or the precise form of duality—primary/auxiliary graphs, low-/high-pass channels, feature/structure streams, homophilic/heterophilic branches, Euclidean/Spherical Harmonics spaces, or primal/dual optimization networks—rather than relying on the label alone.