GNN Training Dynamics MicroSim¶
Run the GNN Training Dynamics MicroSim MicroSim Fullscreen
Edit in the p5.js Editor
About This MicroSim¶
This MicroSim simulates GNN training dynamics on a synthetic node classification task as you vary depth (1–8 layers), dropout rate, and whether residual connections are enabled. Each configuration runs a mini-training loop and plots training loss and validation accuracy curves in real time.
The key phenomenon is over-smoothing: stacking too many layers causes node representations to become indistinguishable, collapsing validation accuracy even as training loss decreases. The gap between the two curves is the over-smoothing signature.
Learning objective (Bloom's Analyze (Level 4)): See how training loss and validation accuracy diverge as GCN depth increases, making the over-smoothing feedback loop concrete. Observe how residual connections and dropout mitigate the collapse.
How to Use¶
- Set depth — drag the layers slider (1–8) to change network depth.
- Toggle residuals — enable or disable residual (skip) connections.
- Set dropout — adjust the dropout rate applied to each layer's output.
- Run training — click "Train" to animate the loss and accuracy curves over 200 epochs.
- Compare runs — previous runs are overlaid in lighter colors so you can compare configurations.
Iframe Embed Code¶
You can embed this MicroSim in any web page with the following HTML:
<iframe src="https://AnvithPothula.github.io/graph-neural-networks-textbook/sims/ch08-gnn-training-dynamics/main.html"
height="532"
width="100%"
scrolling="no"></iframe>
Lesson Plan¶
Grade Level¶
Undergraduate / Graduate (College Level)
Duration¶
20–30 minutes
Prerequisites¶
Gradient descent and backpropagation (Chapter 0). GCN message passing (Chapter 6). Concept of over-smoothing.
Activities¶
- Start with 2 layers (no residuals, no dropout) and train to convergence. Record val accuracy. Increment layers to 4, 6, 8. Plot depth vs. val accuracy.
- At depth 6, enable residual connections. How much does over-smoothing improve?
- At depth 6, apply dropout rate 0.5. Compare to the residual-connection fix — which helps more on this task?
Assessment Question¶
Prove that repeated application of the GCN propagation rule \(D^{-1/2}AD^{-1/2}\) causes node features to converge. What is the limit each feature converges to?
References¶
- Li et al. (2018). Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Classification. AAAI.
- Chen et al. (2020). Simple and Deep Graph Convolutional Networks. ICML.
Part of Chapter 8: GNN Training, Augmentation, and Practical Tips. Return to the chapter page or browse all MicroSims.