Fusing Global Topology and Local Message Passing

20 concepts16 papers1 proposal

Standard Graph Neural Networks (GNNs) operate primarily via message passing, where nodes iteratively aggregate information from their immediate neighbors. While highly effective for capturing local feature correlations, this mechanism has a fundamental limitation: it views the graph through a highly localized lens. As a result, standard GNNs struggle to capture macro-structural properties—such as long-range loops, cycles, or global connectivity patterns—and are theoretically bounded by the Weisfeiler-Leman (WL) graph isomorphism test.

When a network's critical information resides in its global shape rather than immediate neighborhoods (e.g., in sparse infrastructure grids or complex biological networks), message passing alone falls short. To overcome this bottleneck, a prominent research direction focuses on extracting explicit topological and macro-structural descriptors to complement neural embeddings, treating them as distinct, synergistic views of the same data.

The Locality Bottleneck and the Topological Solution

Instead of forcing a neural network to implicitly learn global geometry through deep stacking (which often leads to over-smoothing or over-squashing), modern frameworks explicitly compute structural invariants.

A primary tool for this is Persistent Homology (PH), a method from algebraic topology that tracks how geometric features (like connected components, loops, and voids) appear and disappear across different spatial scales. In On the Expressivity of Persistent Homology in Graph Learning, Rubén Ballester et al. demonstrated that features derived from persistent homology provide highly complementary information to traditional structural features, proving that PH is at least as expressive as the WL hierarchy, and in certain scenarios, even more so.

To make these topological summaries compatible with deep learning architectures, the data must be vectorized without losing its mathematical stability. To address this, Cheng Xin et al. introduced a differentiable 2-parameter persistence vectorization method in GRIL: A $2$-parameter Persistence Based Vectorization for Machine Learning. Their work demonstrates that macro-structural topological summaries provide essential global context that complements the local neighborhood focus of standard message-passing architectures.

Why the Synergy Works: Orthogonal Information Channels

The core intuition behind fusing topology with neural learning lies in the orthogonality of their information channels. Local message passing excels at mapping continuous node features and immediate relational dependencies. In contrast, topological descriptors capture discrete, invariant structural properties that are entirely blind to node features but highly sensitive to global connectivity.

This synergy becomes crucial in sparse or infrastructure-based networks where global loops and paths dictate the system's behavior. In PHLP: Sole Persistent Homology for Link Prediction - Interpretable Feature Extraction, Junwon You et al. showed that integrating topological features from PH with latent GNN embeddings significantly improves link prediction accuracy. Local neighborhood aggregations miss macro-scale connectivity, but global topological structures fill this gap.

This principle extends beyond persistent homology to other forms of macro-structural decomposition:

  • Ensembles of Fundamental Patterns: In Explainable Graph Representation Learning via Graph Pattern Analysis, Xudong Wang et al. showed that decomposing graph representations into a weighted ensemble of fundamental patterns (such as paths, cycles, and cliques) yields superior classification accuracy compared to standard GNNs alone. This suggests that complex graph data is best captured by a diverse ensemble of topological features rather than a single architectural bias.
  • Statistical and Core-Periphery Metrics: Rather than relying on expensive iterative message passing, Raul Barbosa et al. combined local connectivity metrics with global importance and hierarchical core-periphery metrics in A bag of nodes primer on weightless graph classification, providing a fast, multi-faceted view of graph topology.

Multi-View Fusion Mechanisms

To effectively leverage both local neural features and global topological structures, modern architectures treat them as distinct "views" within a unified framework, fusing them through three primary strategies.

1. Contrastive Learning

A powerful paradigm is to align or contrast the local neural view with the global topological view to force the network to encode higher-order shapes. In TopoGCL: Topological Graph Contrastive Learning, Yuzhou Chen et al. introduced a framework that uses Extended Persistence Landscapes (EPL) and a dedicated topological contrastive loss to learn higher-order structural features that standard GNNs overlook. Similarly, Jianqing Liang et al. developed GNN-Transformer Cooperative Architecture for Trustworthy Graph Contrastive Learning, which leverages graph topology (such as k-nearest neighbors based on hop count) directly as a distinct view in contrastive learning to complement node feature representations.

2. Tensor and Parallel Stream Fusion

Other approaches process the views in parallel and fuse them using advanced algebraic operations to preserve multi-scale interactions. Elynn Chen et al. developed Tensor-Fused Multi-view Graph Contrastive Learning, which combines structural graph representations from a GCN with topological features from Extended Persistent Homology using tensor decomposition.

┌──────────────────────────────┐
│         Input Graph          │
└──────────────┬───────────────┘
               │
       ┌───────┴───────┐
       ▼               ▼
┌─────────────┐ ┌─────────────┐
│ Local View  │ │ Global View │
│   (GCN)     │ │ (Topology)  │
└──────┬──────┘ └──────┬──────┘
       │               │
       └───────┬───────┘
               ▼
┌──────────────────────────────┐
│   Tensor / Parallel Fusion   │
└──────────────────────────────┘

This tensor-based architecture style is also mirrored in Bridging Domain Adaptation and Graph Neural Networks: A Tensor-Based Framework for Effective Label Propagation by Tao Wen et al., where combining graph convolutional features with topological information from PH allows for effective domain adaptation. In a non-tensor approach, Qingqing Long et al. designed Graph Structural-topic Neural Network to process node features and structural topic features in parallel streams, combining them via a nonlinear layer so that structural context directly enhances node features.

3. Hierarchical and Subgraph Sampling

Rather than computing global topology on the entire graph at once, some frameworks build intermediate representations to scale efficiently. In Sampling Subgraph Network with Application to Graph Classification, Jinhuan Wang et al. introduced a stochastic sampling strategy ($S^2GN$) that integrates sampled structural features with capsule-based neural networks, capturing complex connectivity patterns more efficiently than standard message passing alone.

Broad Applicability Across Domains

The necessity of balancing local and global views spans multiple specialized domains, proving that the synergy is a fundamental requirement for robust graph modeling:

By treating global topology not as a replacement for message passing, but as an orthogonal, complementary view, these diverse approaches successfully bypass the fundamental expressivity limits of traditional graph neural networks.

Key Foundational Work

Graph Structural-topic Neural Network (2020) — contributed 2 concept elements to this direction.
Sampling Subgraph Network with Application to Graph Classification (2021) — contributed 1 concept elements to this direction.
A bag of nodes primer on weightless graph classification (2021) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_multi_view_graph_representation_108a21cd,
  title = {multi-view graph representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-view-graph-representation/rdr_108a21cd2a2942b1a19a1729b270db9f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_multi_view_graph_representation_108a21cd_v1,
  title = {multi-view graph representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-view-graph-representation/rdr_108a21cd2a2942b1a19a1729b270db9f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake