torch-geometric

To build a Graph Neural Network for node classification, use PyTorch Geometric (PyG) to define the model, format graph edges in edge_index, and run a standard training loop to propagate node features and predict labels.

Handling heterogeneous graph data in PyG uses the HeteroData structure to manage varying node and edge types, enabling you to train models that process complex relational data for tasks like heterogeneous link prediction.

Scaling GNN training for large graphs uses the NeighborLoader to sample localized subgraphs, allowing standard training loops to process massive datasets without exceeding memory limits during node or graph classification.

PyG supports link prediction by providing data structures and model examples that learn edge representations, letting you train Graph Neural Networks to predict missing connections between nodes.

You need PyTorch installed as a prerequisite, because PyG builds directly upon its framework to provide the Data structures, transforms, and model layers required for Graph Neural Network training.