Hands-On Graph Neural Networks: Node Classification and Link Prediction with PyTorch Geometric

This is an open-source graph neural network project (gnn-node-link-pytorch) for beginners, built on the PyTorch Geometric toolkit. It corely supports two main tasks: node classification and link prediction, implements mainstream models like GCN, GraphSAGE, and GAT, and is validated on classic datasets Cora and PubMed to help users quickly get started with GNN technology.

Graph Neural Networks (GNNs) are specifically designed to handle graph-structured data such as social networks and molecular structures, but traditional deep learning frameworks lack native support for them. PyTorch Geometric (PyG) supplements PyTorch with a GNN toolkit, lowering the barrier to building and training GNN models.

Node Classification: Predict the category of unlabeled nodes using known node labels, e.g., predicting user interests in social networks; Link Prediction: Predict whether a connection exists between nodes, e.g., predicting product preferences in recommendation systems or protein-protein interactions.

The project supports three types of models:

• GCN: Aggregates neighbor node features to update representations; it is a foundational work in GNNs;
• GraphSAGE: Samples neighbors and aggregates features, supporting large-scale graphs and inductive learning;
• GAT: Introduces an attention mechanism to assign weights to different neighbors, focusing on key connections.

Datasets:

• Cora: 2708 ML papers (7 categories, 1433-dimensional bag-of-words features, citation edges);
• PubMed: 19717 diabetes papers (3 categories, testing model scalability); Tools: Integrates GNNExplainer to identify key subgraphs and node features that affect predictions, improving model interpretability.

1. Select a dataset (Cora/PubMed or custom);
2. Configure parameters (learning rate, number of training epochs, etc., default values provided);
3. Run training (automatic training and validation);
4. View results (prediction results and performance metrics).

GNNs are applied in multiple fields:

• Recommendation systems: Analyze user-item interaction graphs to capture collaborative signals;
• Drug discovery: Predict properties and interactions using molecular graphs;
• Social network analysis: Community identification, influential user discovery;
• Knowledge graph completion: Predict missing relationships to support intelligent Q&A.

The project provides a friendly entry point for GNN learning, supporting multiple models and datasets. It is recommended that developers further explore: details of attention mechanisms, optimization of large-scale graph training, cross-domain application cases. GNN technology will play a greater role in more practical scenarios in the future.