GNN-Exp: A Visual Exploration Tool for Graph Neural Networks in Computational Notebooks

GNN-Exp is an interactive visualization library designed specifically for graph neural networks, developed by researchers at the University of Minnesota. It supports direct exploration of graph structures, model architectures, and intermediate feature representations in Jupyter Notebooks, offering multiple visualization modes such as node-link views, matrix views, and subgraph sampling. This tool aims to address the understanding and debugging challenges posed by the "black-box" nature of GNN models. It is compatible with mainstream GNN models built using PyTorch Geometric, helping users intuitively analyze graph data and model behavior.

Graph Neural Networks (GNNs) have become the mainstream method for processing graph-structured data, with wide applications in social network analysis, molecular property prediction, knowledge graph reasoning, and other fields. However, the "black-box" nature of GNN models poses challenges for model understanding and debugging—users find it difficult to intuitively grasp how models process graph structure information and how feature representations evolve across layers. Existing deep learning visualization tools are mostly designed for traditional neural networks and struggle to present the unique topological relationships of graph structures; while tools specifically for GNNs often have limited functionality and cannot cover the complete analysis process from raw graph data to internal model computations. The GNN-Exp project was created to address this pain point, providing a comprehensive GNN visualization solution that supports interactive exploration in computational notebook environments.

GNN-Exp includes three main function modules:

GraphVisualizer: Explore raw graph data structures and properties, supporting node-link graphs and adjacency matrix views. It provides subgraph sampling based on "cycle count" to focus on local neighborhood analysis.

GraphEditor: Allows modification of graph structures (adding/removing nodes/edges, modifying properties) within notebooks, and can export to JSON format, facilitating exploration of how graph structure changes affect model behavior.

GNNVisualizer: Delve into the model to display feature transformation processes across layers, supporting tasks like node classification and edge prediction. It allows layer-by-layer viewing of intermediate activation values, and for large graphs, supports subgraph feature visualization to optimize performance.

The frontend of GNN-Exp is developed using TypeScript and React, integrated with the Python backend via the AnyWidget framework, balancing modern browser graphics capabilities with Python ecosystem integration.

It supports three rendering backends: WebGPU (priority), WebGL (fallback), and SVG (static export). Users can specify the backend via the renderer parameter. The frontend is built using esbuild, with hot module reloading supported for development; the uv tool is recommended for managing Python dependencies.

GNN-Exp currently supports various mainstream GNN layer types, including GCNConv, GATConv, SAGEConv/GraphSAGEConv, GINConv, Linear layers, and activation functions like ReLU and Tanh. Its broad architecture support makes it compatible with most GNN models built using PyTorch Geometric—developers only need to pass in the model and data to get an interactive analysis interface.

Typical use cases for GNN-Exp include:

Model Debugging and Understanding: Locating feature propagation issues in abnormal samples; Teaching Demonstrations: Real-time display of graph convolution principles and feature evolution; Paper Figure Generation: Outputting high-quality vector visualizations for academic illustrations; Hyperparameter Tuning Assistance: Observing changes in feature distribution to guide tuning directions.

GNN-Exp has been released on PyPI (pip install gnn-exp) and is open-source under the MIT license. The release process is automated via GitHub Actions, using PyPI trusted publishing to ensure security.

Current limitations: The visualization of attention coefficients for GATConv does not yet support precise per-edge and per-head display. Future directions include enhancing the visualization capabilities for attention mechanisms and promoting such tools to become standard in GNN development and deployment.