GraphTopoNet: Topology-Enhanced Graph Learning

• GraphTopoNet is a research family of models that augment graph neural networks with higher-order topology, using persistent homology and simplicial complexes.
• These models employ techniques such as simplicial message passing, local PH rewiring, and witness-complex pooling to enhance expressivity and robustness.
• GraphTopoNet variants have demonstrated improved performance in diverse tasks including graph classification, node classification, and geophysical mapping.

GraphTopoNet denotes a class of topology-enhanced graph-learning models that augment graph neural computation with higher-order structure, persistent-homological summaries, or other topology-derived signals. In the formulation most directly grounded in topological neural networks, a graph $G=(V,E)$ is lifted to a simplicial complex, simplicial message passing is performed over $k$-simplices, persistence diagrams are computed from filtrations such as vertex-color, edge-color, or geometric $i$-simplex-color filtrations, and the resulting descriptors are injected into simplex features or graph-level readouts (Verma et al., 2024). In adjacent usage, the term also covers architectures that use local persistent homology for graph rewiring, topology-aware pooling, or localized node- and edge-level descriptors, and it also appears as the proper name of a spatial GCN for Greenland bed topography mapping with uncertainty-aware training (Chen et al., 2021, Chen et al., 2023, Yan et al., 15 Jan 2025, Tama et al., 10 Sep 2025). This suggests that GraphTopoNet is best understood as a research family rather than a single canonical network.

1. Terminological scope and research lineage

The most explicit graph-topological formulation comes from the TopNets framework, where a GraphTopoNet instance starts from a graph, constructs a simplicial complex such as the clique complex, applies simplicial message passing over nodes, edges, triangles, and higher simplices, computes persistent homology from chosen filtrations, vectorizes the resulting persistence diagrams, and reinjects those vectors into simplex features and readouts. In that setting, GraphTopoNet is a graph-specialized realization of “Topological Persistent Neural Networks” and unifies or subsumes methods such as TOGL, PersLay, and RePHINE (Verma et al., 2024).

A second usage arises in papers that do not name their method GraphTopoNet in the title but explicitly present their models as topology-enhanced graph neural architectures of that kind. TRI-GNN is described as a prototypical GraphTopoNet because it computes persistent homology on local node neighborhoods, rewires the graph using Wasserstein distances between persistence diagrams, and injects topology-derived weights into message passing (Chen et al., 2021). Likewise, Wit-TopoPool is presented as informing a “GraphTopoNet” built from local topological pooling and witness complex-based global embeddings, while PEGN is presented as a GraphTopoNet-style architecture that augments node classification and link prediction with localized extended persistent homology (Chen et al., 2023, Yan et al., 15 Jan 2025).

A third usage is terminologically narrower. In Greenland bed mapping, GraphTopoNet is the actual model name of a 3-layer GCN that operates on a 4-neighbor grid graph, concatenates raw surface observables with gradient features and polynomial trend surfaces, and couples radar confidence weighting, BedMachine fallback supervision, and Monte Carlo-dropout-based epistemic uncertainty regularization (Tama et al., 10 Sep 2025).

A recurrent misconception is therefore that GraphTopoNet denotes one fixed architecture. The literature instead supports a broader view: the shared denominator is topology-aware graph learning, while the operational mechanisms vary substantially across simplicial lifting, local PH rewiring, witness-complex pooling, localized EPH features, tensorized persistent-image learning, and uncertainty-aware spatial GCNs. The term should also be distinguished from “TopoNet,” which is a separate end-to-end framework for topology reasoning in driving scenes rather than a PH-based graph-learning family (Li et al., 2023).

2. Core architectural patterns

In the TopNets-based graph instantiation, the underlying object is an abstract simplicial complex $K$ over a vertex set $V$, closed under inclusion. Given a graph $G=(V,E)$, the default lift is the clique complex $\mathrm{Cl}(G)$, whose simplices are all cliques in the graph. Message passing then proceeds dimension-wise over faces, cofaces, and upper adjacencies. A generic update with PH injection is

$$x^{(t+1)}_{\sigma} = \phi^{(t)}\!\left( x^{(t)}_{\sigma}, \bigoplus_{\tau \in \mathrm{cofaces}(\sigma)} \psi^{(t)}(x^{(t)}_{\tau}), \bigoplus_{\tau \in \mathrm{faces}(\sigma)} {\psi'}^{(t)}(x^{(t)}_{\tau}), r^{(t)}_{\sigma} \right),$$

where $r^{(t)}_{\sigma}$ is a PH-derived vector for simplex $\sigma$. A TopNets layer first refines simplex features with a TNN layer, then computes persistence diagrams, vectorizes them, performs topological aggregation through mappings such as TopAgg, and finally forms inter-layer and inter-dimension readouts (Verma et al., 2024).

TRI-GNN follows a different computational grammar. For each node $k$0, it extracts a $k$1-hop induced subgraph $k$2, computes persistence diagrams under degree-based or attribute-based filtrations, and defines a topological distance

$k$3

These distances drive Topology-induced Multigraph Representation (TIMR), which adds edges between topologically similar neighborhoods and removes edges between topologically dissimilar ones. The same distances also enter STAN, a topology-aware aggregation mechanism in which neighbors are weighted by softmax-normalized inverse persistence-diagram distances before being consumed by a GNN layer built over the topology-induced Laplacian $k$4 (Chen et al., 2021).

Wit-TopoPool organizes topology through pooling rather than rewiring. It computes node-local persistence diagrams on learned similarity neighborhoods, scores nodes by aggregated lifespans, selects top-$k$5 nodes, and pairs this with a global witness complex branch built on a landmark subset. The witness persistence diagram is vectorized into a witness persistence image, transformed by an MLP, and fused with the pooled representation. The resulting architecture explicitly separates a local topological pooling branch from a global topological embedding branch (Chen et al., 2023).

PEGN and related localized-PH models operate at node and edge-pair scale. For node classification, each node $k$6 is associated with a $k$7-hop vicinity graph $k$8; for link prediction, each candidate pair $k$9 is associated with a pair-vicinity graph $i$0. Extended persistence diagrams are computed on those local complexes, vectorized as persistence images or PI$i$1, and then used either to gate message passing, to concatenate topology into node updates, or to decode link scores (Yan et al., 15 Jan 2025).

In the Greenland formulation, GraphTopoNet abandons PH entirely and instead uses a straightforward stack of three GCNConv layers: $i$2 Its graph topology is spatial rather than algebraic-topological: nodes are grid cells, edges connect 4-neighbors, and topology enters through the graph structure plus uncertainty-aware supervision rather than through persistence diagrams (Tama et al., 10 Sep 2025).

3. Persistent-homological signal construction

Persistent homology enters GraphTopoNet-style models through filtrations. In TopNets, a filtration $i$3 is a nested sequence of subcomplexes, often induced by a filtering function on vertices or simplices. For a vertex-color filtration,

$i$4

which ensures that faces do not appear later than cofaces. The same framework also allows edge-color filtrations and geometric $i$5-simplex-color filtrations. In the geometric case,

$i$6

where $i$7 is permutation-invariant and $i$8 is built from $i$9-invariant geometric quantities such as distances, angles, or volumes (Verma et al., 2024).

The output of a filtration is a persistence diagram $K$0, which records birth and death times of topological features. GraphTopoNet variants use several vectorizations. PersLay aggregates transformed diagram points by an order-invariant operator. Persistence images use Gaussian smoothing: $K$1 TopNets also mention landscapes, weighted kernels, Betti curves, and learned point transforms $K$2 implemented by MLPs (Verma et al., 2024).

TRI-GNN uses persistent diagrams differently. Rather than vectorizing them into learned embeddings, it keeps them in diagram space and computes Wasserstein distances

$K$3

with $K$4 in experiments. These distances become the primitives for rewiring thresholds, STAN aggregation weights, and the construction of the topology-induced adjacency (Chen et al., 2021).

Wit-TopoPool introduces witness complexes as a more scalable alternative to Vietoris–Rips complexes for global topology. A weak witness complex is built over a landmark set $K$5, and its persistence diagram is vectorized as a witness persistence image. This produces a graph-level topological tensor that complements the node-selection scores derived from local VR-based persistence diagrams on learned similarity neighborhoods (Chen et al., 2023).

Localized EPH models enrich the filtration design further. PEGN uses distance-based filters on $K$6-hop ego-graphs or pair-vicinity graphs, optionally with Ollivier–Ricci-curvature-based edge weights and spectral Heat Kernel Signatures. Ordinary and descending filtrations are combined into extended persistence so that loop deaths are also recorded. PI$K$7 augments persistence images with counts of level-$K$8 nodes, intra-level edges, and inter-level edges, making explicit the structural quantities encoded by the localized diagrams (Yan et al., 15 Jan 2025).

4. Expressivity, stability, and equivariance

The central theoretical claim of TopNets is that persistent homology can strictly increase expressivity. The paper proves the proposition “SWL + PH $K$9 SWL”: there exist pairs of non-isomorphic clique complexes that the Simplicial Weisfeiler–Leman test cannot distinguish, but persistence diagrams from color-based filtrations can. The implication stated in the paper is that a GraphTopoNet fusing simplicial message passing with PH descriptors is strictly more expressive than SWL alone and, by extension, surpasses 1-WL-power GNNs when using clique complexes (Verma et al., 2024).

Localized EPH models make an even more granular expressivity claim. The localized EPD around a node or node pair encodes the triangle count incident to a node, its clustering coefficient, the number of level-$V$0 nodes, the number of edges among level-$V$1 nodes, the number of edges crossing levels $V$2 and $V$3, the persistence diagram of each $V$4-hop neighborhood, and the shortest cycle incident to the node. The paper further states that with shortest-path distance filters, EPDs can distinguish some pairs beyond 3-WL, and with pair-distance filters they can distinguish certain pairs beyond 4-WL under specific filters (Yan et al., 15 Jan 2025).

Stability is another recurring theme. TopNets invoke classical bottleneck and Wasserstein stability, writing

$V$5

with $V$6 depending on the Lipschitz constant of the filtering functions. Because vectorizations such as PersLay, persistence images, and landscapes are stable embeddings, the resulting descriptors inherit stability to small perturbations in inputs, features, or coordinates. The same paper also gives discretization-error bounds comparing continuous and discrete variants of TOGL- and RePHINE-style constructions, explaining why continuous TopNets can improve performance without guaranteeing exact recovery of the ODE solution under weak assumptions (Verma et al., 2024).

TRI-GNN derives a different stability statement. Let $V$7 be TIMR graphs of $V$8 with average degrees $V$9. The theorem states that there exists $G=(V,E)$0 such that

$G=(V,E)$1

where $G=(V,E)$2 is the local $G=(V,E)$3-distance defined through persistent-homological neighborhood comparisons. The paper also formulates a conjecture controlling the change in algebraic connectivity under attribute-based TIMR, linking PH-based rewiring to spectral robustness (Chen et al., 2021).

For geometric tasks, TopNets prove that persistence diagrams from geometric $G=(V,E)$4-simplex-color filtrations are $G=(V,E)$5-invariant. E-TopNets then combine those invariant diagrams with EMPSN-based equivariant message passing and coordinate updates, so that geometric symmetries are handled simultaneously at the level of neural dynamics and topological descriptors (Verma et al., 2024).

A related theoretical strand appears in tensor-view topological graph learning. TTG-NN represents multi-filtration persistence images and graph-convolution outputs as tensors and passes them through a Tensor Transformation Layer with Tucker, CP, or tensor-train parameterization. The paper derives high-probability bounds on in-sample and out-of-sample MSE, with stochastic error

$G=(V,E)$6

and core-estimation error

$G=(V,E)$7

thereby tying low-rank tensor structure to improved sample efficiency (Wen et al., 2024).

5. Empirical regimes and reported performance

The empirical record of GraphTopoNet-style models spans graph classification, node classification, link prediction, molecular property prediction, antibody design, molecular dynamics, and geophysical reconstruction. In the TopNets paper, continuous and $G=(V,E)$8-equivariant variants achieve strong performance across antibody design, molecular dynamics simulation, and drug property prediction, while discrete-to-continuous upgrades improve several graph-classification backbones (Verma et al., 2024). TRI-GNN reports that it outperforms all 14 state-of-the-art baselines on 6 out of 7 graphs and yields up to $G=(V,E)$9 better performance under noisy scenarios (Chen et al., 2021). Wit-TopoPool reports that it significantly outperforms 18 baseline models across 11 graph-classification datasets (Chen et al., 2023). PEGN reports state-of-the-art performance on various node-classification and link-prediction benchmarks (Yan et al., 15 Jan 2025). The Greenland GraphTopoNet reports up to $\mathrm{Cl}(G)$0 error reduction relative to interpolation and convolutional baselines on three subregions (Tama et al., 10 Sep 2025).

Variant	Task	Selected reported result
Continuous GraphTopoNet with GIN + RePHINE	Graph classification	NCI1 $\mathrm{Cl}(G)$1; IMDB-B $\mathrm{Cl}(G)$2
Continuous GraphTopoNet with GCN + RePHINE	Graph classification	IMDB-B $\mathrm{Cl}(G)$3
TRI-GNN$\mathrm{Cl}(G)$4	Node classification	Cora-ML $\mathrm{Cl}(G)$5; PubMed $\mathrm{Cl}(G)$6
Wit-TopoPool	Graph classification	MUTAG $\mathrm{Cl}(G)$7; REDDIT-B $\mathrm{Cl}(G)$8
PEGN	Link prediction	Cora $\mathrm{Cl}(G)$9 AUC-ROC; PubMed $$x^{(t+1)}_{\sigma} = \phi^{(t)}\!\left( x^{(t)}_{\sigma}, \bigoplus_{\tau \in \mathrm{cofaces}(\sigma)} \psi^{(t)}(x^{(t)}_{\tau}), \bigoplus_{\tau \in \mathrm{faces}(\sigma)} {\psi'}^{(t)}(x^{(t)}_{\tau}), r^{(t)}_{\sigma} \right),$$0 AUC-ROC
Greenland GraphTopoNet	Bed topography mapping	Upernavik RMSE $$x^{(t+1)}_{\sigma} = \phi^{(t)}\!\left( x^{(t)}_{\sigma}, \bigoplus_{\tau \in \mathrm{cofaces}(\sigma)} \psi^{(t)}(x^{(t)}_{\tau}), \bigoplus_{\tau \in \mathrm{faces}(\sigma)} {\psi'}^{(t)}(x^{(t)}_{\tau}), r^{(t)}_{\sigma} \right),$$1; Hayes RMSE $$x^{(t+1)}_{\sigma} = \phi^{(t)}\!\left( x^{(t)}_{\sigma}, \bigoplus_{\tau \in \mathrm{cofaces}(\sigma)} \psi^{(t)}(x^{(t)}_{\tau}), \bigoplus_{\tau \in \mathrm{faces}(\sigma)} {\psi'}^{(t)}(x^{(t)}_{\tau}), r^{(t)}_{\sigma} \right),$$2; Kangerlussuaq RMSE $$x^{(t+1)}_{\sigma} = \phi^{(t)}\!\left( x^{(t)}_{\sigma}, \bigoplus_{\tau \in \mathrm{cofaces}(\sigma)} \psi^{(t)}(x^{(t)}_{\tau}), \bigoplus_{\tau \in \mathrm{faces}(\sigma)} {\psi'}^{(t)}(x^{(t)}_{\tau}), r^{(t)}_{\sigma} \right),$$3

The reported gains are not uniform in mechanism, but they do align around a common empirical pattern: augmenting graph learning with topology helps when the target depends on cycles, motifs, higher-order connectivity, shape robustness, or symmetry. In TopNets, higher-order PH and continuous dynamics improve IMDB-B and PROTEINS and support geometric tasks such as QM9, CDR-H3 antibody design, and 3BPA MD17. In TRI-GNN, both TIMR and STAN are necessary; removing either degrades performance substantially, and the rewired topology improves robustness under random edge attacks (Verma et al., 2024, Chen et al., 2021).

The Greenland use case is qualitatively different but still informative for the name’s applied scope. Nodes are $$x^{(t+1)}_{\sigma} = \phi^{(t)}\!\left( x^{(t)}_{\sigma}, \bigoplus_{\tau \in \mathrm{cofaces}(\sigma)} \psi^{(t)}(x^{(t)}_{\tau}), \bigoplus_{\tau \in \mathrm{faces}(\sigma)} {\psi'}^{(t)}(x^{(t)}_{\tau}), r^{(t)}_{\sigma} \right),$$4 grid cells at $$x^{(t+1)}_{\sigma} = \phi^{(t)}\!\left( x^{(t)}_{\sigma}, \bigoplus_{\tau \in \mathrm{cofaces}(\sigma)} \psi^{(t)}(x^{(t)}_{\tau}), \bigoplus_{\tau \in \mathrm{faces}(\sigma)} {\psi'}^{(t)}(x^{(t)}_{\tau}), r^{(t)}_{\sigma} \right),$$5 m resolution, features combine surface elevation, velocity components, surface mass balance, thickening rate, finite-difference gradients, and degree-2 polynomial trend surfaces, and uncertainty is estimated with $$x^{(t+1)}_{\sigma} = \phi^{(t)}\!\left( x^{(t)}_{\sigma}, \bigoplus_{\tau \in \mathrm{cofaces}(\sigma)} \psi^{(t)}(x^{(t)}_{\tau}), \bigoplus_{\tau \in \mathrm{faces}(\sigma)} {\psi'}^{(t)}(x^{(t)}_{\tau}), r^{(t)}_{\sigma} \right),$$6 Monte Carlo-dropout forward passes. Reported performance includes Upernavik MAE $$x^{(t+1)}_{\sigma} = \phi^{(t)}\!\left( x^{(t)}_{\sigma}, \bigoplus_{\tau \in \mathrm{cofaces}(\sigma)} \psi^{(t)}(x^{(t)}_{\tau}), \bigoplus_{\tau \in \mathrm{faces}(\sigma)} {\psi'}^{(t)}(x^{(t)}_{\tau}), r^{(t)}_{\sigma} \right),$$7, RMSE $$x^{(t+1)}_{\sigma} = \phi^{(t)}\!\left( x^{(t)}_{\sigma}, \bigoplus_{\tau \in \mathrm{cofaces}(\sigma)} \psi^{(t)}(x^{(t)}_{\tau}), \bigoplus_{\tau \in \mathrm{faces}(\sigma)} {\psi'}^{(t)}(x^{(t)}_{\tau}), r^{(t)}_{\sigma} \right),$$8, $$x^{(t+1)}_{\sigma} = \phi^{(t)}\!\left( x^{(t)}_{\sigma}, \bigoplus_{\tau \in \mathrm{cofaces}(\sigma)} \psi^{(t)}(x^{(t)}_{\tau}), \bigoplus_{\tau \in \mathrm{faces}(\sigma)} {\psi'}^{(t)}(x^{(t)}_{\tau}), r^{(t)}_{\sigma} \right),$$9, SSIM $r^{(t)}_{\sigma}$0, and PSNR $r^{(t)}_{\sigma}$1; Hayes MAE $r^{(t)}_{\sigma}$2, RMSE $r^{(t)}_{\sigma}$3, $r^{(t)}_{\sigma}$4, and SSIM $r^{(t)}_{\sigma}$5; and Kangerlussuaq MAE $r^{(t)}_{\sigma}$6, RMSE $r^{(t)}_{\sigma}$7, $r^{(t)}_{\sigma}$8, and SSIM $r^{(t)}_{\sigma}$9 (Tama et al., 10 Sep 2025).

6. Limitations, misconceptions, and open directions

The main limitation across PH-centric GraphTopoNet variants is computational. Higher-order message passing, clique-complex construction, witness-complex construction, and persistence-diagram computation add overhead; TopNets explicitly note worst-case $\sigma$0 PH complexity, TRI-GNN gives neighborhood-comparison complexity on the order of $\sigma$1, and localized EPH approaches emphasize the cost of recomputing diagrams during end-to-end learning (Verma et al., 2024, Chen et al., 2021, Yan et al., 15 Jan 2025). These costs motivate truncated complexes, selected-layer PH computation, sparse data structures, union–find for $\sigma$2-dimensional persistence, and witness complexes instead of full Vietoris–Rips constructions.

A second limitation is design sensitivity. Filtration choice strongly affects performance in TopNets and PEGN; threshold choices $\sigma$3 matter in TRI-GNN; landmark selection matters in Wit-TopoPool, where degree and betweenness landmarks outperform random selection; and continuous formulations require solver tuning to avoid instabilities (Verma et al., 2024, Chen et al., 2021, Chen et al., 2023, Yan et al., 15 Jan 2025). In the Greenland model, dependence on BedMachine fallback supervision is identified as a limitation because it can propagate BedMachine biases, especially in slow-flowing regions (Tama et al., 10 Sep 2025).

A third misconception is that “topology-aware graph learning” necessarily means persistent homology. The literature is broader. TTG-NN uses PH but emphasizes tensorized multi-view fusion and low-rank tensor transformations (Wen et al., 2024). Topological communities define topology through a rich vector of node properties, PCA, and clustering in principal-component space, yielding role-based, non-contiguous communities that complement geometric communities. That framework is not a PH-augmented message-passing model, but it demonstrates another rigorous route by which topology can structure graph learning and graph summarization (Seoane, 2024).

The future directions stated in these works are relatively consistent. TopNets point to more scalable PH, richer complexes such as cellular complexes, stronger continuous-time theory, solver-aware training, and tighter integration with tensor-field or SE(3)-style representations. TRI-GNN points to scalable PH approximations, multi-filtration strategies, and mitigation of societal bias concerns in neighborhood-based aggregation. PEGN points to richer topology beyond the $\sigma$4-skeleton, adaptive neighborhoods, and scalable differentiable PH operators. The Greenland GraphTopoNet points to distributed training, hierarchical graph constructs, and explicit physics constraints such as mass conservation (Verma et al., 2024, Chen et al., 2021, Yan et al., 15 Jan 2025, Tama et al., 10 Sep 2025).

Taken together, these developments define GraphTopoNet not as a single model class with a fixed layer equation, but as a technically diverse research program: one that uses topology—persistent, simplicial, witness-based, localized, tensorized, role-based, or spatial—to expand the representational repertoire of graph learning beyond ordinary pairwise message passing.