Week 03 — Graph Neural Networks

Message passing as the core operation. Why GNNs are permutation-equivariant by construction.

Lecture

Message-passing framework (Gilmer et al. 2017) · update / aggregate / readout · permutation equivariance and invariance · graph convolution as a special case · the canonical layer.

Read before the lecture

• Gilmer et al., *Neural Message Passing for Quantum Chemistry* (ICML 2017)

Code lab

Your first GNN

Train a 2-layer GCN on the Cora citation network using PyTorch Geometric. Compare against a logistic-regression baseline.

Notebook: lab01-gnn-cora.ipynb  ·  Dataset: Cora (citation network, public).

Reference text for this week: chapter 03 of the bilingual notes — EN PDF · FR PDF.