benedekrozemberczki/CapsGNN

Combines capsule networks with graph neural networks to generate multiple graph-level embeddings that capture diverse structural properties, moving beyond scalar node representations. Uses stacked GCN layers with routing mechanisms and attention modules to aggregate node capsules into graph-level capsules, enabling the model to handle variable-sized graphs. Integrates PyTorch Geometric for graph operations and trains end-to-end on graph classification tasks using JSON-formatted graph inputs.

1,280 stars. No commits in the last 6 months.