Graph Neural Networks
Uncover Hidden Structure

How GNNs learn representations from network topology — turning raw citation links into embeddings that reveal community structure raw features miss.

▼

The Network

Class Distribution

Degree Distribution

Neighborhood Label Consistency by Class

Model Showdown

Three GNN architectures compete — each brings a different inductive bias to learning node representations.

Detailed Comparison

Metric	GCN	GAT	GraphSAGE

Embedding Space

Watch how GNN-learned representations separate communities compared to raw features. Toggle between models to see the difference.

Why This Matters

Raw features show overlapping, noisy clusters — the bag-of-words vectors don't capture relational information.
GNN embeddings form tight, well-separated clusters because they aggregate neighborhood information through message passing.
The NMI improvement from 0.13 → 0.59 shows GNNs extract 4.5x more community signal than features alone.

Training Dynamics

Loss curves, accuracy trajectories, and convergence behavior across architectures.

Training Loss

Validation Accuracy

Validation Macro F1

Train vs Validation Accuracy

Community Detection

GNN embeddings vs. traditional Louvain — who finds better communities?

Clustering Quality (NMI vs Ground Truth)

Interactive Network

Explore the network — hover over nodes to see their connections and properties. Force-directed layout reveals natural community structure.

Per-Class Performance

How well does each model distinguish between research areas?

Key Insights

What we learned about GNNs and representation learning on social networks.

Graph Neural NetworksUncover Hidden Structure