Graph Neural Network (GNN)

A graph neural network (GNN) is a deep learning model that learns representations of nodes, edges or whole graphs by repeatedly aggregating information from each node's neighbors.

그래프 신경망(GNN)은 각 노드가 이웃 노드의 정보를 반복적으로 집계하도록 하여 노드·엣지·그래프 전체의 표현을 학습하는 딥러닝 모형이다.

How do I make predictions from data where entities are connected in a network?

개체들이 네트워크로 연결된 데이터에서 어떻게 예측하나요?

A graph neural network learns from both node features and the connections between nodes by passing and aggregating messages along edges. It fits node classification, link prediction and graph classification in social, citation, molecular or transaction networks.

When to use Graph Neural Network (GNN)

• Node features plus an edge list describing connections
• Node classification, link prediction or graph classification
• Large networks where structure carries predictive information

When to use something else

• Explaining tie formation with interpretable parameters; use exponential random graph models
• Tabular data without meaningful connections; use gradient boosting or a multilayer perceptron

Assumptions and requirements

• A defined graph structure (nodes and edges)
• Labelled nodes, edges or graphs for supervised tasks
• Feature scaling and train/validation/test splits that respect the graph
• Sufficient data to avoid overfitting

What the analysis reports

• Accuracy, F1 or AUC on held-out nodes or edges
• Learned node embeddings
• Training and validation loss curves
• Embedding visualizations

How to report it

The graph neural network achieved a test accuracy of [value]% on node classification (mean over [value] runs, SD [value]).

Related methods

• Exponential Random Graph Models
• Node2Vec
• Multi Layer Perceptron

Run Graph Neural Network (GNN) on Quorant