torch-geometric

Build graph neural networks for node classification in PyTorch Geometric by applying message passing layers like GCN, GAT, or GraphSAGE to torch_geometric data structures. You construct custom layers and feed graph data directly into the model for training.

Train large-scale graph neural networks on graphs that do not fit in memory by using neighbor sampling and link sampling loaders. These mini-batch loaders subset local graph neighborhoods, allowing you to process massive graphs efficiently without loading the entire topology.

PyTorch Geometric supports heterogeneous graphs with multiple node and edge types. You can convert relational data into typed graph pipelines and use specialized message passing layers designed to operate across distinct node and edge categories.

Explain graph neural network predictions and inspect model behavior using built-in explanation utilities in PyTorch Geometric. These tools help identify which nodes, edges, or features contribute to specific predictions for model transparency.

Convert relational data into a heterogeneous graph pipeline by mapping distinct entities to typed nodes and relationships to typed edges within torch_geometric data structures. This allows you to apply custom datasets and heterogeneous message passing layers directly.

PyTorch Geometric supports link prediction tasks on custom datasets by utilizing message passing layers and neighbor loaders. You can train models to predict edges between nodes using the same data structures used for node and graph classification.