ml-pytorch-geometric

To build a graph neural network for node classification in PyTorch Geometric, structure your data as a PyG Data object with node features, edge_index, and targets, then apply message passing layers like GCNConv to propagate node embeddings across the graph topology.

Heterogeneous graphs in PyTorch Geometric are handled using the HeteroData construct, which organizes distinct node and edge types into separate feature dictionaries while applying dedicated message passing modules to route and aggregate type-specific graph relationships.

Neighbor sampling for large graph neural networks uses NeighborLoader to extract localized subgraphs around target nodes during training, enabling scalable mini-batch processing and distributed out-of-core learning without loading the full sparse adjacency matrix into memory.

Yes, you can use PyTorch Lightning to manage distributed training for PyTorch Geometric models by wrapping datasets with LightningDataset or LightningNodeData, which streamline training loops and integrate scalable neighbor sampling directly into Lightning datamodules.

PyTorch Geometric models require correct tensor shapes for edge_index and batch vectors because the framework uses these specific dimensions to properly route messages, aggregate sparse adjacency data, and combine multiple disjoint graphs into a single batched representation.

Message passing layers on extremely sparse graph data face memory bottlenecks when scaling to large graphs, requiring neighbor sampling and distributed out-of-core concepts to prevent excessive memory consumption during aggregate and update operations on massive sparse adjacency matrices.