Back to Projects

Deconstructing GNNs from Scratch

Explore the mathematics of message passing and graph convolutions — built entirely from scratch in pure PyTorch.

Part 1

The Math of Message Passing

Breaking down graph convolution, adjacency normalization, neighborhood aggregation, and the GCN spectral perspective.

Part 2

PyTorch Implementation

Building GCN and GAT layers with attention-weighted aggregation — entirely from scratch without torch_geometric.
View Code on GitHub

Part 3

Classifying Nodes in Graphs

Node classification on Zachary's Karate Club with only 4 labeled examples, and visualizing learned attention weights.