Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cross-View Topology-Aware Graph Representation Learning

Ahmet Sami Korkmaz, Selim Coskunuzer, Md Joshem Uddin

TL;DR

<3-5 sentence high-level summary> GraphTCL addresses the limitation of GNNs in capturing global graph topology by introducing a dual-view framework that learns both structural embeddings from a GNN and topological embeddings from persistent homology. These two modalities are explicitly aligned in a shared latent space via a cross-view contrastive loss, while a simple linear classifier yields the final predictions. Empirical results on TU and OGB molecular benchmarks show consistent improvements over strong baselines, with ablations highlighting the key role of cross-view alignment and robustness to filtration choices. The work suggests topology-aware contrastive learning as a principled and practical approach to enhance graph representations without adding architectural complexity.

Abstract

Graph classification has gained significant attention due to its applications in chemistry, social networks, and bioinformatics. While Graph Neural Networks (GNNs) effectively capture local structural patterns, they often overlook global topological features that are critical for robust representation learning. In this work, we propose GraphTCL, a dual-view contrastive learning framework that integrates structural embeddings from GNNs with topological embeddings derived from persistent homology. By aligning these complementary views through a cross-view contrastive loss, our method enhances representation quality and improves classification performance. Extensive experiments on benchmark datasets, including TU and OGB molecular graphs, demonstrate that GraphTCL consistently outperforms state-of-the-art baselines. This study highlights the importance of topology-aware contrastive learning for advancing graph representation methods.

Cross-View Topology-Aware Graph Representation Learning

TL;DR

<3-5 sentence high-level summary> GraphTCL addresses the limitation of GNNs in capturing global graph topology by introducing a dual-view framework that learns both structural embeddings from a GNN and topological embeddings from persistent homology. These two modalities are explicitly aligned in a shared latent space via a cross-view contrastive loss, while a simple linear classifier yields the final predictions. Empirical results on TU and OGB molecular benchmarks show consistent improvements over strong baselines, with ablations highlighting the key role of cross-view alignment and robustness to filtration choices. The work suggests topology-aware contrastive learning as a principled and practical approach to enhance graph representations without adding architectural complexity.

Abstract

Graph classification has gained significant attention due to its applications in chemistry, social networks, and bioinformatics. While Graph Neural Networks (GNNs) effectively capture local structural patterns, they often overlook global topological features that are critical for robust representation learning. In this work, we propose GraphTCL, a dual-view contrastive learning framework that integrates structural embeddings from GNNs with topological embeddings derived from persistent homology. By aligning these complementary views through a cross-view contrastive loss, our method enhances representation quality and improves classification performance. Extensive experiments on benchmark datasets, including TU and OGB molecular graphs, demonstrate that GraphTCL consistently outperforms state-of-the-art baselines. This study highlights the importance of topology-aware contrastive learning for advancing graph representation methods.

Paper Structure

This paper contains 19 sections, 7 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Graph Representation and Contrastive Learning
  4. Topological Methods and Topology Aware Contrastive Learning
  5. Position of GraphTCL.
  6. Background
  7. Graph Neural Networks
  8. Persistent Homology (PH)
  9. Heat Kernel Signature (HKS) as a Filtration Function.
  10. GraphTCL: Topological Contrastive Learning
  11. Contrastive Loss.
  12. Joint Training.
  13. Experiments
  14. Datasets
  15. Implementation Details
  16. ...and 4 more sections

Figures (2)

  • Figure 1: Overview of the GraphTCL framework. Given an input graph, the structural branch encodes node features through a GNN and global pooling, while the topological branch extracts topological features via persistent homology, followed by an MLP encoder. The two embeddings are aligned through a cross–view contrastive loss and jointly optimized with the classification objective.
  • Figure 2: Relative accuracy drop when replacing the HKS-based filtration with degree- or closeness-based filtrations across seven TU datasets