Graph ML Task Types: Node, Edge, and Graph-Level Prediction
You'll learn: How graph machine learning problems are categorized into three fundamental prediction tasks based on what you're trying to predict.
Core Idea
Just as we classify images or predict house prices in traditional ML, graph neural networks solve prediction problems — but on graph-structured data. These problems fall into three natural categories based on what you're predicting:
Node-Level Prediction (Node Classification)
Predict properties of individual nodes using their features and graph structure.
Analogy: In a social network, predicting which users are likely interested in a topic based on their profile and connections. Each node (person) gets a label or value.
Examples: Classifying research papers by field, detecting fraudulent accounts, or predicting protein functions in biological networks.
Edge-Level Prediction (Link Prediction)
Predict whether edges should exist between node pairs, or classify existing edges.
Analogy: Recommending potential friendships on social media — will two users likely connect? You're predicting missing or future edges.
Examples: Product recommendations (user-item edges), predicting drug interactions, or forecasting citation links between papers.
Graph-Level Prediction (Graph Classification)
Predict properties of entire graphs as single units.
Analogy: Given molecular structures (each is a graph), predict whether the molecule is toxic. You treat the whole graph as one instance.
Examples: Classifying molecules by properties, predicting whether a social network community is active or inactive, or identifying malicious code from program dependency graphs.
Why This Matters
These three task types require different architectural approaches and evaluation strategies. Node tasks aggregate neighborhood information, edge tasks reason about node pairs, and graph tasks pool information across the entire structure.
Key Takeaway: Graph ML organizes around three prediction levels — nodes (classify vertices), edges (predict connections), and graphs (characterize entire structures) — each requiring tailored modeling approaches.