Abstract
Local structural information can increase the adaptability of graph convolutional networks to large graphs with heterogeneous topology. Existing methods only use relatively simplistic topological information like node degrees. This paper introduces the Persistence Enhanced Graph Network (PEGN), a novel architecture that leverages persistent homology to capture complex local structures. By quantifying the connectivity and "loopiness" of a node's neighborhood, PEGN learns to intelligently reweight messages during convolution, significantly outperforming standard GNNs on large, dense, and complex datasets.
Introduction
Standard spatial Graph Neural Networks (GNNs), such as GCN or GraphSAGE, update node representations by aggregating messages from neighbors. However, these models are often "topology-blind," relying on simple metrics like node degree to weight messages. Node degree is a shallow metric; it fails to distinguish whether a node is a critical "bridge" in a tree-like structure or a redundant member of a highly connected clique.
For instance, in a social network, a message from a unique link to a different community should be weighted differently than one from a member of a tight-knit group. PEGN addresses this by using mathematically robust topological patterns to guide the deep learning pipeline.
Methods: Persistent Homology and PEGN
To capture these structural nuances, PEGN employs Persistent Homology (PH), a tool from algebraic topology. PH creates a structural "fingerprint" of a node's neighborhood by tracking how topological features—such as connected components and cycles—appear and disappear as a filtration function is applied across the graph.
Since raw persistence diagrams are not directly compatible with neural networks, they are converted into Persistence Images. These are stable, vectorized representations that can be processed by the Persistence Image Network (PIN), a sub-network that outputs a reweighting vector.
Unlike standard GNNs that use scalar weights, PEGN performs channel-wise reweighting. The output of the PIN allows the network to emphasize different features of a neighbor's message based on the local topology. This data-driven approach allows the model to learn the relationship between local structure and feature importance during training.
Experimental Evaluation
PEGN was evaluated on several graph benchmarks, including citation networks (Cora, Citeseer, PubMed), co-authorship networks (Coauthor-CS, Physics), and e-commerce networks (Amazon-Computers, Photo). The model demonstrated significant improvements, particularly on larger and denser datasets where structural nuances are more critical:
- Amazon Computers: Achieved 86.3% accuracy, significantly outperforming GAT (78.0%) and GCN (82.6%).
- Coauthor Physics: Reached 94.2% accuracy, setting a new baseline.
- PubMed: Achieved 79.4% accuracy, surpassing GCN and GraphSAGE.
The results confirm that advanced topological information provides a much richer signal for graph learning than simple node-feature-based attention or node degrees.
Conclusion
PEGN introduces a robust way to incorporate topological stability and complexity into GNNs. While calculating persistent homology introduces some computational overhead, the gains in accuracy on dense and complex graphs are substantial. Future work involves optimizing filtration choices and extending the framework to global graph classification tasks.
Summary updated