📄 Explainable_GNNs.pdf
The slide deck covers:
- What graphs really mean in learning systems
- How GNNs are constructed from graphs (step-by-step)
- Message passing explained using intuition and analogies
- Why classical graph theory is not replaced but operationalized
- Why explainability is essential for graph-based learning
- A structured taxonomy of GNN explainers
- Brain network examples and motivation from neuroscience
➡️ The slides are intuition-first, with minimal mathematics and heavy use of analogies, visuals, and conceptual grounding.
All notebooks are:
- ✅ Google Colab–safe
- ✅ CPU-only
- ✅ Tested with compatible package versions
- ✅ Beginner-friendly, with extensive comments
Notebook_A_GNN_Basics.ipynb
What you’ll learn:
- How a graph is represented as tensors
- What node features, edge indices, and labels mean
- How message passing works computationally
- How graph structure becomes learnable information
Audience:
Ideal for mathematicians and graph theorists seeing GNNs for the first time.
Notebook_B_GNN_Training.ipynb
What you’ll learn:
- How a simple GCN is defined
- How node classification works
- What loss functions mean in graph settings
- Why training GNNs is different from CNNs / RNNs
Includes:
- Karate Club dataset
- Clear separation of model, data, and training loop
Notebook_C_GNN_Explainability.ipynb
This is the heart of the repository.
Covers:
- GNNExplainer (subgraph-based explanations)
- Feature attribution on graphs
- Node vs edge importance
- Visualization of explanations
- Why explanations are local and model-specific
Key takeaway:
An explanation is not “why the graph is like this”,
but “why the model behaved this way on this graph”.
🎯 https://neurograph-fmri-gnn-explorer-695890547372.us-west1.run.app/
A lightweight demo application that:
- Loads a trained GNN
- Runs inference on a graph
- Visualizes important nodes and edges
- Allows users to see explanations instead of just numbers
Designed for:
- Live demos
- Classroom use
- Workshops and interdisciplinary talks
In graph-based domains such as:
- Brain connectomes
- Social networks
- Biological interaction networks
- Infrastructure graphs
Predictions alone are not sufficient.
We must ask:
- Which nodes mattered?
- Which connections mattered?
- Is the explanation stable?
- Does it align with domain knowledge?
This repository treats explainability as a scientific requirement, not a cosmetic add-on.
This material is especially suitable for:
- Mathematicians curious about ML on graphs
- Graph theorists exploring modern learning paradigms
- Researchers working on brain and biological networks
- Faculty teaching AI to interdisciplinary audiences
- Students seeking intuition before equations
No prior deep learning expertise is required.
- Start with the slides to build intuition
- Run Notebook A to understand how graphs enter neural networks
- Proceed to Notebook B to see learning in action
- Explore Notebook C to understand explainability
- Use the demo app for visualization and engagement
- Kipf & Welling, Semi-Supervised Classification with Graph Convolutional Networks
- Gilmer et al., Neural Message Passing for Quantum Chemistry
- Ying et al., GNNExplainer: Generating Explanations for Graph Neural Networks
- Battaglia et al., Relational Inductive Biases, Deep Learning, and Graph Networks
Ebin Deni Raj
FACTS-H Lab
Indian Institute of Information Technology Kottayam
Focus areas:
- Responsible & Explainable AI
- Graph-based learning
- Interdisciplinary AI education
- Trustworthy machine learning systems
If this repository helps you:
- understand GNNs better,
- explain them more clearly,
- or trust them a little more,
then it has served its purpose.
Feel free to fork, reuse, and adapt for teaching and research.