Graph Neural Networks Learning Path

Graph representation and permutation properties

Graph neural networks (GNNs) are designed for graph-structured inputs, where nodes typically have no canonical ordering. To handle this, GNN layers are built to be permutation equivariant: if you reorder the nodes in the input graph, the corresponding node representations are reordered in the same way. For graph-level predictions, GNNs use permutation-invariant readout (global pooling) so the final output does not change under any node/edge ordering. A central mechanism in GNNs is message passing, where each node iteratively updates its representation by aggregating information from its neighbors. This message-passing formulation is expressed using a permutation-invariant aggregation operator (e.g., sum/mean/max), ensuring the update does not depend on neighbor ordering. Architecturally, GNNs often combine permutation-equivariant message-passing layers, local pooling (to coarsen the graph and expand receptive field), and permutation-invariant global pooling (to produce fixed-size graph embeddings).

0/2 modes complete

Permutation equivariant layers

0/2 modes complete