torch-geometric

To build a node classification pipeline in PyTorch Geometric, construct a `Data` object with `edge_index` and use layers like `GCNConv` to train on graph-structured data. You implement the forward pass, loss calculation, and evaluation loop to classify nodes.

The `edge_index` format for graph learning tasks is a tensor of shape [2, num_edges] defining source and target nodes. Correct `edge_index` conventions are required for `Data` and `HeteroData` objects to properly propagate features during message passing.

Yes, you can handle heterogeneous graphs in PyG using the `HeteroData` object. It stores separate feature matrices and edge indices for different node and edge types, allowing architectures like HGT or RGCN to learn heterogeneous graph representations.

Neighbor sampling for scalable mini-batch training of Graph Neural Networks samples multi-hop neighborhoods around target nodes to construct manageable subgraphs. This prevents loading the entire graph into memory, enabling efficient training on large-scale graphs.

Yes, PyTorch Geometric supports link prediction and graph classification tasks. For link prediction, it decodes node embeddings to predict edges, and for graph classification, it pools node representations to classify entire graph structures.

You should implement custom `MessagePassing` layers when built-in GNN architectures like GCN, GAT, or GraphSAGE cannot express your specific graph convolution logic. This allows defining custom message, aggregation, and update functions.