← All resources

GNN From Scratch

Auto-discovered

Blog Intermediate

curlysaarthak

Visit resource ↗

Description

GNN From Scratch is a Notion technical note by curlysaarthak that rebuilds the classical graph neural network model with math intuition and from-scratch code.

It covers graph and adjacency-list representation, the transition function and message passing, state convergence via contraction mappings and fixed-point ideas, the output function, a full GNN architecture, and training with total loss plus Rprop. Snippets show neighbor influence and node forcing networks, with pointers to the 2009 Scarselli et al. paper and Distill or Hugging Face intros.

Readers need basic machine learning and the usual prerequisite math. The author treats modern GNN stacks as advanced but presents the original formulation as approachable.

The page is a free public Notion document with headings, formulas, and code. Further reading includes Stanford CS224W lecture pointers.

It is useful as an end-to-end sketch of classical GNN message passing and convergence before adopting newer graph libraries.