Part 1 - Foundations
Start here if you only know basic ML (vectors, matrices, neural nets).
- Graph Foundations
What is a graph, adjacency matrix, Laplacian, graph tasks, why graphs beat flat ML on relational data. - Representation & Data Preparation
Node/edge/graph features, building graphs from raw data, splits, leakage, PyTorch Geometric, DGL.
Part 2 - Core GNN Theory
The heart of graph deep learning: message passing and classic architectures.
- Message Passing Framework
MPNN, permutation invariance/equivariance, aggregation, update - the unified view behind all GNNs. - Classic Architectures
GCN, GraphSAGE, GAT, GIN, MPNN - full math, intuition, code, and when to pick each. - Spectral Methods
Graph Fourier transform, ChebNet, spectral vs spatial convolution - where GCN really comes from.
Part 3 - Advanced Architectures
Going deeper: limitations, heterogeneous graphs, time, and transformers on graphs.
- Deep GNNs & Limitations
Oversmoothing, over-squashing, expressiveness, Jumping Knowledge, residual connections, normalization. - Heterogeneous & Knowledge Graphs
R-GCN, HGT, metapaths, link prediction, knowledge graph embedding + GNN hybrids. - Temporal & Dynamic Graphs
Discrete snapshots, continuous-time (TGN), evolving graphs, temporal link prediction. - Graph Transformers
Graphormer, GPS, structure-aware attention, positional encodings on graphs.
Part 4 - Practice & Frontier
Training at scale, choosing GNNs, LLM integration, and latest research.
- Training, Optimization & Scaling
Losses, sampling, mini-batch, full-graph, distributed training, frameworks, debugging. - When to Use GNNs & Evaluation
Decision framework, metrics, benchmarks, GNN vs MLP vs Transformer - honest tradeoffs. - GNNs in the LLM Era
Text-attributed graphs, GraphRAG, hybrid LLM+GNN systems, when graphs still matter. - Recent Advances & Papers
2020–2026 breakthroughs, foundation models on graphs, surveys, thesis pointers, code repos.
Quick Reference
Formula cheat sheet: GNN Formula Sheet · Key papers: Papers Hub