Lesson 46 • Advanced
Graph Neural Networks (GNNs) 🕸️
Learn how GNNs process graph-structured data for social networks, drug discovery, fraud detection, and knowledge graphs.
What You'll Learn in This Lesson
- • How message passing works in graph neural networks
- • Node classification for fraud detection in social networks
- • Graph-level prediction for molecular property estimation
- • Real-world GNN applications across industries
- • Aggregation, normalization, and readout functions
🎯 Real-World Analogy: Imagine you're at a party and want to know if someone is trustworthy. You don't just look at them — you look at who their friends are, and who their friends' friends are. That's exactly what GNNs do: each node learns from its network neighbourhood.
Try It: Message Passing
Watch how nodes aggregate neighbour information through a GNN layer
Python
import numpy as np
# ============================================
# GRAPH NEURAL NETWORKS: MESSAGE PASSING
# ============================================
np.random.seed(42)
print("=== How GNNs Work: Message Passing ===")
print()
print("In a graph, each node learns from its neighbours.")
print("It's like gossip: each person updates their opinion based")
print("on what their friends think, then shares their new view.")
print()
# Define a small social network
# 0 -- 1 -- 2
# | |
#
...Try It: Fraud Detection with GNN
Classify fraudulent users by propagating signals through a social network
Python
import numpy as np
# ============================================
# NODE CLASSIFICATION WITH GNN
# ============================================
np.random.seed(42)
print("=== Node Classification: Fraud Detection in Social Network ===")
print()
print("Task: Some users in a network are fraudsters. Can we detect")
print("them by looking at their connections AND features?")
print()
# 8-node network with known and unknown labels
# fraud=1, legit=0, unknown=-1
n_nodes = 8
labels_true = [0, 0, 1, -1,
...Try It: Molecular Property Prediction
Use GNNs to predict drug-like properties from molecular structure
Python
import numpy as np
# ============================================
# GNNs FOR MOLECULAR PROPERTY PREDICTION
# ============================================
np.random.seed(42)
print("=== Molecular Graph: Predicting Drug Properties ===")
print()
print("Molecules are naturally graphs: atoms = nodes, bonds = edges.")
print("GNNs predict properties like toxicity, solubility, binding affinity.")
print()
# Represent a simple molecule: Aspirin (simplified)
# C-C(=O)-O-C(=O)-C1=CC=CC=C1-O
atoms = ["C1",
...📋 Quick Reference — GNN Variants
| Model | Aggregation | Best For |
|---|---|---|
| GCN | Mean of neighbours | Node classification |
| GAT | Attention-weighted | Heterogeneous graphs |
| GraphSAGE | Sampled neighbours | Large-scale graphs |
| GIN | Sum + MLP | Graph classification |
| MPNN | General framework | Molecular property |
🎉 Lesson Complete!
You understand GNNs! Next, learn how AutoML automates model selection and architecture search.
Sign up for free to track which lessons you've completed and get learning reminders.