Graph Neural Networks — Complete Guide

GNNs learn from graph-structured data — social networks, molecules, knowledge graphs.

Why GNNs?

Traditional ML: Assumes independent data points
Graph data: Data points are CONNECTED

Examples:
├─ Social networks: Users connected by friendships
├─ Molecules: Atoms connected by bonds
├─ Citations: Papers connected by references
└─ Knowledge graphs: Entities connected by relations

Message Passing

GNN Layer:

For each node:
1. Collect messages from neighbors
2. Aggregate messages
3. Update node representation

h_v^(t+1) = UPDATE(h_v^(t), AGGREGATE({h_u^(t) : u ∈ N(v)}))

Where:
h_v = node v's representation
N(v) = neighbors of v

GNN Variants

GCN (Graph Convolutional Network):
├─ Spectral-based
├─ Averaging neighbor features
└─ Simple and effective

GraphSAGE:
├─ Sample neighbors (scalable)
├─ Multiple aggregation functions
└─ Inductive (works on new nodes)

GAT (Graph Attention Network):
├─ Attention over neighbors
├─ Learn which neighbors matter
└─ More expressive than GCN

Key Takeaways

1. GNNs learn from graph-structured data
2. Message passing aggregates neighbor information
3. GCN is the simplest GNN
4. GraphSAGE scales to large graphs
5. GAT learns attention over neighbors
6. Applications: molecular property prediction, social networks, recommendations
7. Over-smoothing is a challenge for deep GNNs
8. Graph Transformers combine GNNs with attention